Review the attached article titled : "Information and Password Attacks on Social Networks: An Argument for Cryptography". Then use the Article Review Template to complete the written review. Must b

Information and Password Attacks on Social Networks:: An Argu ment for

Cryptography

Article   in  Journal of Information Technology Rese arch · July 2017

DOI: 10.4018/ JITR.2015010103

CITATIONS

14

READS

877

3 author s:

Some of the author s of this public ation are also w orking on these r elated pr ojects:

OwlBe ans Vie w pr oject

Coding Vie w pr oject

Enric o Franchi

Univ ersit à di P arma

26 PUBLICA TIONS    201 CITATIONS    

SEE PROFILE

Ag ostino P oggi

Universit à di P arma

193 PUBLICA TIONS    4,630 CITATIONS    

SEE PROFILE

Michele T omaiuolo

Universit à di P arma

89 PUBLICA TIONS    469 CITATIONS    

SEE PROFILE

All c ontent f ollowing this p age w as uplo aded by Michele T omaiuolo on 03 August 2015.

The user has r equested enhanc ement of the do wnloaded file. Information Attacks on Online Social Networks

Enrico Franchi, Agostino Poggi and Michele Tomaiuolo

Department of Information Engineering, University of Parma, Italy

ABSTRACT

Online social networks have changed the way people interact, allowing them to stay in touch with their

acquaintances, reconnect with old friends, and establish new relationships with other people based on

hobbies, interests, and friendship circles. Unfortunately, the regrettable concurrence of the users’

carefree attitude in sharing information, the often sub-par security measures from the part of the system

operators and, eventually, the high value of the published information make online social networks an

interesting target for crackers and scammers alike. The information contained can be used to trigger

attacks to even more sensible targets and the ultimate goal of sociability shared by the users allows

sophisticated forms of social engineering inside the system. This work reviews some typical social

attacks that are conducted on social networking systems, carrying real-world examples of such

violations and analysing in particular the weakness of password mechanisms. It then presents some

solutions that could improve the overall security of the systems.

Keywords: social networks; information security; cryptography; password cracking; social engineering.

INTRODUCTION

If we have to chose among the innovations of the past decade just a single phenomenon because of its

outstanding social impact, that would be the diffusion of online social networks. While some social

networking services were already active in the nineties, the capillary diffusion and the sheer number of

people involved transformed online social networking in an unprecedented revolution only recently.

Online social networks have already modified the way people interact. They allow users to reconnect

with old friends, or to establish new connections with unknown people. In general, users are facilitated

in maintaining and developing the relationships with their acquaintances, on the basis of common

activities, interests, and contacts.

From a technological perspective, online social networks are mostly based on sets of web-based

services that allow people to present themselves through a profile, to establish connections with other

users in the system, and to publish resources. Moreover, these systems use common interests and the

natural transitivity of some human relationships to suggest new contacts with whom to establish a

connection. Although some of these aspects already appeared in other systems, online social networks

represent an unprecedented cultural phenomenon. It is mainly characterised by the unceasing flow of

information that users publish in such systems, and their relentless strife to increase the number of their

virtual friends and acquaintances.

Unfortunately, online social networks are becoming an interesting target for crackers and scammers

alike. In fact, many factors concur to attract malicious actions: ( i ) the users’ carefree attitude in sharing

information, ( ii ) the often sub-par security measures from the part of the system operators, and ( iii ) the

high value of the published information. In particular, the information available through social media

can be used to trigger attacks to even more sensible targets. Various forms of social engineering inside

the system, including sophisticated and long term attacks, may be facilitated by the general sense of

sociability shared by users, which per se is an intrinsic objective of social media. However, while the only lasting solution to privacy and security issues would be increasing the users awareness, much can

and shall be done at the system level in order to protect the data with cryptography and to decrease the

impact of wrong choices and mistakes on the user’s part.

This work reviews some typical social attacks that are conducted on social networking systems,

carrying real-world examples of such violations and analysing in particular the weakness of password

mechanisms. It then presents some solutions that could improve the overall security of the systems.

SECURITY THREATS ASSOCIATED WITH SOCIAL MEDIA

Nowadays, online social networks involve people from the entire world, of any age and with any kind

of education. They also helped to increase computer usage among categories that previously showed

little interest for it (Stroud, 2008). The users of information systems have various types of security

requirements, including: confidentiality , integrity , accountability , availability and anonymity . The same

security requirements can be applied to social networking platforms, as well.

Unfortunately, while most users are aware that their profile and the information they publish is

essentially public, they usually strengthen their privacy settings only after problems arise and tend to

overlook the actual impact of the information they disclose (Stroud, 2008). Apparently harmless

information can be exploited, and the more information the attacker has, the more severe and

sophisticated the attack can be. For example, name, location and age can be used to connect a profile to

a real-world identity for more than half of the residents in the USA (Irani et al., 2011).

In fact, social networking platforms are susceptible to different types of attacks, targeting different

components, conducted from different domains, using different techniques. For better analysing these

attacks, it is useful to identify the main abstract components of a generic social networking platform,

corresponding to different functional aspects of those systems. Attackers can target each of the different

components, or they can target different levels, possibly with roughly the same logic. We identify four

main components:

1. The social networking component. It manages and protects access to the users' personal profiles

and the social relationships among users.

2. The content management component. It manages and protects access to all user generated

content, including personal status updates, comments, links to other content, photos and

multimedia galleries.

3. The infrastructure services component. It provides the basic infrastructure services needed to

run the social networking platform, including storage and replication services for content and

profiles, information indexing and routing, management of users' online presence.

4. The communication and transport component. It encapsulates basic inter-networking and ad-

hoc networking functionalities.

Moreover, we can distinguish two different kinds of attackers:

1. Intruders. An attack can be conducted by users accessing the system without proper

authorization, or with accounts created for conducting the attack, purposely.

2. Insiders. Also legitimate users or entities participating in the systems operations can assume

malicious behaviours. From the users point of view, malicious behaviour can also be attributed

to the service provider.

All the various classical kinds of security attacks may be adapted and applied to online social networks,

too. The most typical threats against OSNs include: • Unauthorized Access. Users who have not been granted adequate permissions for accessing

some services and resources, may attempt to circumvent the security mechanisms and policies

of the system and gain unauthorised access. In a social networking platform, any user who has

access to some profiles and messages can harm their legitimate owners. The collection of

existing data is the basis of profiling attacks. These data may also supply some knowledge for

secondary data collection from a wide range of different sources, including other OSNs.

Remote access can also occur at system level. In this case the attacker may directly gain control

of all resources.

• Social engineering. In a social networking application, a common attack is to psychologically

manipulate a user into performing misguided actions. It is similar to a confidence trick or a

traditional fraud, but by means of computer-based communications and online social

networking, typically to gain access to confidential information. In a phishing scheme, the

attacker masquerades as a trustworthy entity to obtain the desired information. In most cases the

victim and the attacker never acknowledged each other directly in real life.

• Masquerading. When a rogue user disguises his identity and claims the identity of another user,

the former is said to be masquerading. Masquerading may be attempted by an attacker either

during a conversation or while registering his own profile, for deceiving other users or the

whole social networking platform. Sometimes, masquerading is the first step to gain access to

infrastructure services and resources to which the attacker is not entitled. Simple impersonation,

by cloning the victim's profile from the same platform or by porting profile data from a different

platform, may easily lead the attacker to gain trust from the victim's contacts. This way, it can

damage other users eventually deceived. Especially in communities where reputation is valued,

masquerading can also damage the user whose identity has been stolen. In fact, the attacker may

pretend to be another user in order to shift the blame for any liable action.

We will provide some examples and a discussion of these threats in the following subsections.

However, also other traditional security attacks can be conducted against OSNs, including:

• Denial of Service (DoS). The services and communications at the infrastructure level can be

disrupted by common denial of service attacks. Social networking platforms are also susceptible

to all the conventional denial of service attacks aimed at the underlying operating system or

communication protocols. In addition to attacking the whole infrastructure of a social

networking platform, users can also launch denial of service attacks against specific users,

especially in a distributed platform. For example, repeatedly sending messages or other spam

may place undue burden on the recipient users and their systems. Malicious users can also

intentionally distribute false or useless information to prevent other users from completing their

social activities.

• Repudiation. In general, repudiation occurs when a user, after having performed some action,

later denies that action having happened (at least under his responsibility). Repudiation can be

intentional or even accidental. It can also be the result of a misunderstanding, when users have a

different view of events. In any case it can generate important disputes. In a sense, nothing can

prevent a user from repudiating one of his actions. But a social networking platform can

eventually help resolving disputes by providing needed evidence, if it maintains a sufficiently

detailed log of events. For users who value their reputation, the availability of such evidence

may constitute a valid deterrent. • Eavesdropping. The attempt to observe the flow and possibly the content of confidential

messages is one of the most classical security threats. Apart from reading the content of

messages, which may require cryptanalysis, an eavesdropper may gather useful information by

simply observing the pattern of messages and their recipients, for example inferring the type of

services being requested. To eavesdrop on other users, an attacker may also exploit the

infrastructure and communication services of the platform, e.g. through unauthorized access.

• Alteration. When a user signs up a social networking service, he starts exposing his profile and

content to the platform. An attacker may tamper with the profile and content data published by

the victim, with all the messages he communicates to other users and all data used on the

infrastructure services. Alteration can also be conducted by the service operator, which provides

the facilities for online social networking and may take control of published data. Alteration

may take the particular form of filtering, or censorship, when applied systematically for

removing undesired content from the OSN.

• Copy and Replay. Each action in a social network may be subject to copy and reply. In this type

of security threat, an attacker attempts to intercept some data and clone it, for retransmitting it

later. The interceptor may successfully copy and replay a message, a complete profile or any

other data. If those data are not associated with a signature and a timestamp, the repeated

reception of such copies may pass unnoticed and accepted as a legitimate action.

Types of information leaks from OSNs

According to Li, Li and Venkatasubramanian (2007), there are two types of privacy attacks in online

social networks:

1. Identity disclosure. It occurs when the adversary is able to determine the mapping from an

anonymous profile to a specific real-world entity (e.g. an individual).

2. Attribute disclosure. It occurs when an adversary is able to determine the value of a user

attribute that the user intended to stay private.

Moreover, attackers may gather data from:

1. A single social networking platform.

2. Multiple platforms, services or extension apps.

Sensible data may be obtained both through:

1. Direct access , either through an authorized service or through some breach, and

2. Attribute inference , based on public data available about a user's contacts, and the correlation

usually existing among attributes of linked profiles.

Either about identity or attributes, privacy in online social networks is mostly intended as user-to-user

privacy: even when the relative settings are set correctly, so that no other user in the system can access

information not intended for his eyes, the system itself has full access to information. In fact, in most

online social networks, the system owners actually rely on such information to make a profit, for

example to improve the accuracy of target advertisement. Unfortunately, as long as they have full

access to the information – i.e., the information is available in the system in the clear – any security

issue or naivety results in privacy violations. Even innocuous features can be easily become serious issues. For example, in order to promote the

service, Facebook had a public directory containing the names of the users that used to be presented

along with 10 random friends of theirs. Such feature allowed web-spiders to repeatedly request pages

from such directories and essentially discover the structure of the network. Subsequently, the number of

friends was reduced to 8 and the selection of friends became a deterministic function of the IP address

of the requester. However, even the greatly reduced information can still be used to violate privacy

(Bonneau, 2009).

Another very relevant problem is the amount of integration we expect from digital services. Online

social networks are relatively open system that can be enhanced by third party widgets and games,

usually called apps, which are embedded in the user’s pages and typically have access to some amount

of user information. Moreover, the users often want different services to interoperate for various

reasons, so that, for example, a new tweet is also notified to the Facebook friends or pictures stored on

Flickr are accessed by an online printing service. Eventually, the credentials of the online social

network can be used, in many cases, to log in a different system without needing a separate account.

The problem is that when any of these external services, systems or apps is violated, private

information can be stolen. For example, in June 2012, although the main Twitter servers were not

compromised, sensible data was nonetheless stolen from a third party widget, TweetGif (Robertson,

2012).

Similarly, we expect our mobile devices to interact cleverly with the online social networks. Service

providers create mobile applications, which, in turn, have some degree of access to the data stored on

the mobile device. Although such applications usually do not maliciously access user’s data for

purposes different from those stated, sometimes privacy is neglected. For example, the iPhone

LinkedIn App used to send all the user’s calendars to central LinkedIn servers, including phone

numbers, call details and passcodes, while only the relevant information should have been sent (Cheng,

2012).

Even when the social networking platform does not provide data directly, yet many supposedly private

properties may be inferred about a user. In fact, in social networks the attributes of users who share

some kind of link are often correlated. Zheleva and Getoor (2009) introduce different models of

attacks, to infer the hidden sensitive values on the basis of friendship links or public group membership

data. Specifically:

• Friend-aggregate. It looks at the sensitive attribute distribution amongst the friends of the

person under question.

• Collective classification. It aims at inferring class labels of linked objects together, instead of

classifying each instance independently of the rest. It iteratively uses inferred values for

connected private profiles, in addition to public ones.

• Flat-link. It deals with links by flattening the data and considering the adjacency matrix of the

graph. Each user has a list of features, corresponding to the size of the network. The public

profiles are used as a training set for the classifier.

• Blockmodelling. It is a stochastic model, supposing that users form natural clusters or blocks.

Then, their interactions can be explained by the blocks they belong to.

• Groupmate-link. In this model, groupmates are considered as friends to whom users are

implicitly linked. It assumes that each group is a clique of friends. Thus, it creates a friendship

link between users who belong to at least one group together, without representing the strength

of the link. The resulting network can then be analysed using one of the previous models.

• Group-based classification. It considers each group as a feature in a classifier, and sensitive attributes are inferred from the groups a user belongs to.

• Basic. It calculates the overall marginal distribution of public data and uses it to infer sensitive

attributes of private profiles. This is the simplest model, which can be applied also in the

absence of relationship and group information.

Phishing and impersonation attacks

Social engineering has always been a major security threat to information systems. Two of the more

frequent kinds of social engineering attacks, i.e., pretexting and phishing, require some background

information, but such information was not always easy to obtain. In fact, some of the early social

engineering attacks actually involved going through company or people's trash bins in order to find

scraps of paper containing relevant information. Nowadays, many interesting pieces of information are

publicly available in social networking systems or accessible only to victim's friends at best (or at

worse, from the attacker’s perspective).

Traditional phishing consisted in sending electronic communications to a huge number of individuals

with the intent to have them disclose some relevant information, such as passwords or credit card

numbers. Since the large scale of such attacks, even if only few users actually trusted the

communication and disclosed information, it was nonetheless enough to make the attack profitable.

Such communications typically used very generic ways to address the victim and typically avoided any

specific detail so that the same message could be sent to all the recipients. Nowadays, people have been

instructed not to trust such generic messages, and other forms of phishing were created.

Spear phishing is a form of phishing where the attack is directed against a single individual and the

offending communication contains large amounts of information on the victim himself, in order to

convince him of the communication authenticity. This strategy is very effective both if the information

employed is public but the victim somewhat thinks it is not or if the attackers gained access to private

information by other means. Another popular form of phishing is cloning: a cloning attack requires an

original legitimate message that is subsequently tampered and sent the victims. These communications

look extremely authentic, but nonetheless they contain ways to direct the victims towards the attackers

website (where he is usually requested his credentials). Data-leaks offer an even more dangerous

variant: when the data leak becomes public, users expect their service providers to send them emails

notifying them to change their password. If such mails are instead well-crafted phishing emails, that

perhaps also use some private information about the user, many more people fall for the scam. The kind

of attack just described actually happened in the days immediately after the LinkedIn breach.

A serious weakness in many online services regards the “security questions”. In fact, while phishing

can be avoided by paying attention to the actual identity of the communications initiators, not much can

be done against “security questions” attacks. Some services allow the users to choose the security

questions so that the information required to answer it correctly is not publicly available, however, for

many other sites only a closed list of security questions is available. Researchers estimated that more

than 33% of the security questions involve names of relatives or friends and about 16% involve

personal preferences (“favorite something”). Popular security questions are also the name of a pet

(typically the current or first one), ZIP code or social security number (Rabkin, 2008). The victim in

online social networks often directly or indirectly divulges such pieces of information. Even apparently

safer information, such as social security numbers, may be guessed with high confidence from publicly

available data (Acquisti & Gross, 2009).

This kind of attacks is usually not directed towards online social networks, at least initially, but usually

targets email service providers, because, when the security question is answered correctly, most

systems send a password-reset email, and, consequently, controlling the main email account is

necessarily the first step. Similarly to spear phishing, security-question related attacks target only specific individuals. Popular examples are the violation of Sarah Palin's email account in 2008, using

information publicly available on Wikipedia and on official websites and, more recently, the attack

against Mat Honan. In the latter case, the tech journalist had his mail, iCloud and Twitter accounts

compromised after an elaborate chain of social attacks, during which also his mobile phone, tablet and

laptop have been remotely wiped (Honan, 2012).

Another typical attack on social networking systems is impersonation (identity theft), which can occur

either on a site where the victim already has a profile (profile cloning) or on different site (cross-site

profile cloning). In both cases, it is easy to convince the victim’s acquaintances to accept friendship

from the fake profile and subsequently to disclose confidential data. Experiments on the effectiveness

of impersonation have been conducted and the results show that such attacks mostly succeed (over 50%

success rate) even in the case of simple non-cross-profile cloning (Bilge et al., 2009).

When the attackers obtain credential for an online social network profile, additional attacks can be

performed. If the attackers are interested in keeping a private database containing the violated accounts

data, they can create a widget that copies the relevant resources or simply scrapes the web pages.

Moreover, they can try to use the same credential to attack the user’s accounts in other online services.

Since users do not often use different passwords for different accounts, this strategy is rather effective.

For example, during a period of roughly a week in June 2012, several services were compromised.

About a month after this so-called “breach week”, a number of Dropbox accounts have been hacked,

using the passwords stolen from the already compromised accounts (Agarwal, 2012).

Another possibility is using the violated account to accept friendships from profiles controlled by the

attackers. If the victim has already enough friends (on average, a Facebook user has 1000 friends), it is

not likely that he notices the new friendship, especially because the attackers can remove any

notification before the account owner actually sees that. Such friendship can be subsequently used as a

Trojan horse to access the user's data even after the user changes the password and, moreover, is also

likely to attract friendship from the victim contacts, considering that mutual friendships usually

increases the confidence that an accounts should be trusted.

Controlling many profiles in a social network, either real ones or fake, can be used for spam purposes,

to influence opinions in general or simply to gather huge amounts of information. In fact, in order to

either access the data of most users in the network or deliver them a message appearing to come from a

friend, it is not necessary to control a large number of accounts, but it is sufficient to control accounts

that approximate a dominating set, i.e. a set D such that V = D friends(D)∪ , where V is the set of all

users in the network. Although computing a minimal dominating set is an NP-complete problem and

requires full knowledge of the networks, decent approximations can be made using greedy algorithms

and using only sampled knowledge.

THE WEAKNESS OF PASSWORDS

Most of the attacks we discussed so far are forms of social engineering. However, online social

networks are not, from a technical point of view, different from any other password-protected service:

if the passwords are not chosen judiciously and stored in a safe way, then the violation of accounts

becomes very easy. In fact, the analysis of available records about the robustness of users' passwords,

as well as the security mechanisms and policies deployed in online systems, does not shape a

reassuring scenario. Table 1 summarizes some major password leaks suffered by social websites in

recent years. Date Target site Size of leak Password storage

2012, June LinkedIn 6,5 million passwords SHA1, not salted

2012, June eHarmony 1,5 million passwords MD5, not salted

2012, June last.fm 2,5 million passwords MD5, not salted

2012, June League of Legends 32 million passwords MD5, not salted

2011 PlayStation Network 77 million accounts Hashed, not salted

2011 Sony Online Entertainment 25 million accounts Hashed, not salted

2009 Rockyou 32 million accounts Clear text

Table 1. Major password leaks

The “Breach Week”

The first week of June 2012 has earned the fame of “Breach Week”. Some million passwords were

stolen, in order, from LinkedIn, eHarmony and last.fm accounts. At least the first two incidents appear

correlated, as the passwords were published by the same user on a Russian forum about cracking. The

first file, leaked from the professional social site LinkedIn, totalled 6,5 million unique passwords. It is

not clear if other passwords were leaked, but the episode may have involved many more users. The

second file, leaked from the popular online dating site eHarmony, contained around 1,5 million

passwords. Finally, the internet radio platform last.fm notified all its users of a possible leak of

passwords, requesting to update their login information. Administrators announced ongoing

investigations. A published file accounts for at least 2,5 million passwords leaked form the radio

platform. But these episodes are not isolated. In fact, few days later, an alert appeared on the online

gaming platform League of Legends, notifying its 32 million users of a breach in some of its databases

(Merrill & Beck, 2012). The site administrators revealed that more than half of the passwords would

not resist to crack: “ We compared encrypted password hashes and discovered that 11 passwords were

shared by over 10,000 players each... A double-digit percentage of individuals had the same password

as at least one other person. ” Apart from revealing a widespread lack of awareness among users about

basic password security, the announcement may imply that passwords, like on LinkedIn and eHarmony,

were hashed but not salted (Yin, 2012).

Sony, in 2011, suffered from breaches in its PlayStation Network and Qriocity media streaming service.

The services were suspended for almost a month. Personal data from around 77 million accounts were

stolen, possibly including credit card numbers. During the same period, Sony Online Entertainment

announced that another breach led to leaking personal data from 25 million accounts. These breaches

compare to previous massive data leaks from TJX Companies, a fashion retail network, in 2007,

affecting over 45 million customers, including details about credit card and in some cases also social

security numbers and driver's license numbers. In 2009, Heartland Payment Systems, currently the fifth

largest credit card processor in the United States and the 9th in the world, was affected by a breach

which costed something in the tens of million dollars. In 2005, a similar breach at CardSystems

Solutions, another payment card processor, jeopardized roughly 40 million credit and debit card

accounts (Vijayan, 2007).

One of the best known episodes, finally, is the breach suffered by Rockyou. After an SQL injection

attack in 2009, the intruders gained access to the website database, which included the full list of

unprotected clear text passwords of all 32 million users. In fact, those passwords are still readily available over the BitTorrent network and are being used for dictionary attacks against other websites

(Cubrilovic, 2009).

Common password storing and cracking techniques

The safe management of passwords is an old problem. Various password storing mechanisms can be

considered.

1. Clear text. Though used in practice, the storage of passwords as clear text should out of the

question, as it offers no protection against intruders.

2. Encryption. The use of traditional encryption schemes is also discouraged. In fact, by

knowledge of the decryption key, all passwords may be subverted in a single shot. If an intruder

acquires the control of a machine, then the possibility of loosing a decryption key is quite

concrete.

3. Hashing. The solution adopted since decades in Unix systems is based on cryptographic one-

way functions, that can only be inverted by guessing the original clear text password (Morris &

Thompson, 1979). However, common hashing algorithms are often designed for efficiency,

which allows attackers to try many combinations in short time. Moreover, the effort to guess

users' passwords can be reduced by attackers, if they generate the hash of a tentative password

and confront it with each one of the actual password hashes of the attacked system.

4. Salting. If some unique value (a salt) is added to each password before hashing it, the result is

unique for each user. If two users use the same password, two different hashes are obtained,

since that password is combined with two different salts. Then, in the database, both the hash

and the salt, in the clear, need to be stored. Thus, it is not possible to pre-compute hashes for all

popular and simple passwords, or for all combinations generated through brute force (Morris &

Thompson, 1979).

5. Password hashing algorithms. While common hashing algorithms are designed to be as fast and

efficient as possible, password hashing algorithms are designed to require a significant amount

of computational resources. Bcrypt, one of the best options among password hashing

algorithms, is based on the Blowfish algorithm and allows developers to decide the number of

iterations of its main function, possibly requiring various orders of magnitude more time than

generic hashing algorithms. The exact choice depends on the desired balance of password

security and needed computational resources for normal operation, in particular for handling the

regular number of logins (Provos & Mazieres, 1999).

Though password storing mechanisms are well known and documented, they are not always used in

existing systems, including some popular services, with large user bases. In fact, some lessons can be

learned about implemented mechanisms for password security in real cases.

A number of sites adopts techniques that are far from the best practices in this field. We will leave the

Rockyou case apart. LinkedIn, for example, avoided storing passwords in clear text, but used a

suboptimal algorithm for hashing. In fact, it used a good generic hashing algorithm (SHA-1, namely),

instead of a password hashing algorithm, like bcrypt. On moderate hardware, SHA-1 can be computed

over almost 200MBs of data per second, and MD5 over more than 300 MB of data per second (Dai,

2009). With these algorithms, a password of 6 lowercase alphanumeric characters can be easily

obtained through a brute force attack in less than a minute. And this is without using the potential of

parallel GPU computing, which can obtain results which are at least an order of magnitude better.

Exploiting four HD 5970 cards and some precalculations for the latest steps of MD5, the Whitepixel

tool may achieve 33.1 billions MD5 hash/s, on a system costing 2.700 $ at the end of 2010 (Bevand, 2010).

Another lesson that can be learned is that many websites simply skip password salting, even if it is a

well established technique (Morris & Thompson, 1979). LinkedIn and eHarmony are not isolated

examples, though emblematic given their huge user bases. For example, it took many years and

versions for the popular blogging platform Wordpress to finally add salt to its user passwords, in 2008

at version 2.5.

In all those careless sites, simple attacks can be based on dictionaries of popular passwords, together

with mangling rules to obtain similar and derived passwords. Another possibility is to try all possible

combinations of lowercase letters, uppercase letters, digits and punctuation symbols, in a brute force

attack. Some tools, just like John the Ripper, can apply both attacks on a given list of hashed

passwords. Starting from a dictionary or a combinatorial engine, the obtained password is hashed and

then compared to all available hashes, possibly leading to the discovery of one or more users'

passwords after a single hash operation. The effectiveness of the operation is greatly simplified by the

fact that a single algorithm is applied against all passwords, without salt or additional parameters.

Moreover, if passwords are not salted, the attacks can be made even more effective by calculating in

advance the hashes of all possible passwords, up to a certain length. Obviously, taking into account the

needed disk space, this approach is feasible only for very short passwords. But techniques are available

to trade time for space, thus reducing the needed disk space but requiring more hash calculations at

runtime. Among such techniques, some are based on the so-called rainbow tables . Oechslin (2003)

shows how a previous technique, described by Hellman and refined by Rivest, could be further

improved, halving the number of calculations during cryptanalysis.

Those methods are all based on the iterative calculation of a hash function and a reduction function, in

an alternating sequence, starting from a given password and repeating the cycle some thousands of

times, depending on the desired balance between space and runtime processing time. For a given chain,

only the starting password and the final hash are stored, while intermediate results are discarded. The

number of chains to store depends on the desired success probability in decrypting a given hashed

password. In the original paper, the method is applied to Windows LanManager passwords. With a

space of 1.4GB for rainbow tables (and thanks to the weakness of the old LanManager scheme) a

success rate of 99.9% can be achieved.

Given a certain hash, finding the corresponding password requires finding a rainbow chain in the table.

If the original hash is not found, then one or more cycle of the reduction function and hash function are

applied and then the search is repeated. Finally, when the relevant rainbow chain is found, starting from

the first password in the chain, all calculations are repeated, till the password associated with the

original hash is found.

ANALYSING SOME MILLION USERS' PASSWORDS

Unfortunately, users do not seem to choose passwords with care. A first glimpse of this fact comes from

the security alert issued after the League of Legends intrusion (Merrill & Beck, 2012). In that alert, the

system administrators pointed out that a high percentage of users had the same password of another

user. The reported data of 10000 different users sharing the same 11 passwords, is a glaring evidence.

So, in order to confirm the impression that users choose very weak passwords, we decided to analyse

the publicly available list of leaked hashes from LinkedIn. Figure 1. Dependence of the number cracked by using both dictionary (labeled as “Basic”, “Full” and

“RockYou”) and brute force attacks, as functions of cracking time. In the result of the “RockYou”

approach it is particularly evident how the “brute-force” attack starts after exhausting the

“dictionary” attack.

The first step is to crack the hashes to obtain the original passwords. Since the hashing procedure is

essentially irreversible, the typical strategy is to “guess” which string generated the hash. One

possibility is to brute-force attack the hash, i.e., simply hashing every possible string and comparing it

with the desired hash. Such strategy is very slow and ineffective in practice whereas dictionary attacks

are much more effective. In a dictionary attack the strings are not chosen at random, but are extracted

from a dictionary of common words. Additionally, the words from the dictionary are mangled

simulating the usual strategies users employ to make passwords harder to guess (e.g., substitute

numbers for alphabetic characters, appending values to the end or the beginning of the words).

Although more techniques exist, we decided to start cracking the passwords using the popular

password cracker “John the Ripper”, which performs both brute force and dictionary attacks with

mangling rules and we reported our results in Figure 1. All the attempts were performed on a Core2

Duo laptop at 2.40 MHz. Our first attempt (“Basic” in Figure 1) used a small dictionary of around 3500

common password that is distributed with John itself and we also enabled simple mangling rules for

generating common derivative passwords, in two hours the program was able to recover more than a

million (1.05 M) LinkedIn users’ passwords. The second attempt (“Full” in Figure 1) used a much

larger dictionary, containing a few million words from more than 20 languages, also available from the

program website. The larger dictionary provides some marginal gain.

Eventually, we fed John with the database of 32 million passwords that were stolen from the RockYou website in 2009 and subsequently published on the web. The results significantly improved. In two

hours around 1.5 million passwords were revealed, 44% more than those obtained with the first

attempts. In this case, the first million passwords were obtained in less than five minutes.

The most likely explanation is that the LinkedIn password set has a large overlap with the RockYou

dataset, which, in turn, confirms the idea that many users tend to reuse passwords and are not extremely

creative in deciding their own passwords. Moreover, in this case, of the 2 million passwords we

cracked in 24 hours starting with the RockYou password dictionary, more than 25% are not longer than

6 characters and that less than 15% are longer than 8 characters. Moreover, almost the 30% consists

only of alphabetic lowercase characters and another 30% consists of alphanumeric characters, with no

special characters, and, eventually, almost 1 password over 10 consists of only digits. Roughly 12% of

the passwords end with the digit 1, 200.000 passwords end with 123 and about half as many with 1234.

Years are also frequently included (e.g., numbers from 1975 to 2011, while 2012 is significantly less

popular and “future” years are rare). The most popular base word that constitutes the password is a

variation of the name of the website (“linked”, “linkedin” and “link”) and “love” and “password” still

made it in the top 10. Analyses of other leaks of passwords show that variations on the service name are

popular and that “password” is an evergreen (Dragusin, 2012).

In fact, the passwords that have not been cracked during our 24h timeframe are expected to be much

stronger (otherwise they would have been cracked as easily) and the fact that only 1/3 of the password

was cracked is somewhat encouraging. It is also interesting that, using the RockYou dataset as

dictionary, the number of password cracked ( N(p) ) depends on the length of the passwords ( L(p) )

according to N(p) = A e -BL(p)

where A = 3.15x106 and B = -0.786 .

The last consideration regards the nature of the dataset: repeated passwords were not included, i.e., in

the original system more than one user could have the same passwords, so that 2 million passwords

could account for far more than 2 million of users, and, considering the remarks of the operators of

League of Legends, they probably do.

HOW TO MAKE ONLINE SOCIAL NETWORKS MORE SECURE?

The huge amount of information and the features-over-security approach of online social networks is a

serious security problem, especially considering that some million of vulnerable accounts remain a big

security threat for other systems as well, in the sense that, considering high password reuse, the more

services are violated, the more likely it is that accounts on other systems are violated as well.

Many solutions can be used to increase security. While the best long-term solution to the security

problems is increasing the users’ awareness, this may be unpractical because of the huge amount of

people involved and their perception to be in a relatively safe environment, where friends are just few

clicks away. However, there are many solutions that providers and software developers can use in order

to improve the overall security.

For, example, most browsers and email clients offer some features to automatically detect phishing and

one of the major weaknesses, i.e., security questions, can be fixed either improving the quality of the

questions themselves (Rabkin, 2008) or, better, sending reset messages via SMS, which at least would

require the attacker to physically access the victim’s mobile phone.

Zhou, Pei and Luk (2008) suggest some anonymization techniques to prevent privacy attacks in social

networks. In particular, they suggest two different approaches, based respectively on generalization and

perturbation:

1. Clustering. Vertices and edges are clustered into groups, so that a subgraph can be abstracted

into a supervertex. This way, details about individuals are hidden.

2. Graph modification. Some vertices and edges of the graph are modified, by insertion or deletion. Such modifications may be applied using an optimization approach, randomly, or

greedily, to match privacy requirements.

Felt & Evans (2008) suggest to protect data of social networks, especially from exposure to third-party

developers. They present a privacy-by-proxy design, which mask data presented to external

applications, using placeholders instead of actual values. However, it is not clear how resistant this

simple approach is against de-anonymization techniques.

Li et al. (2007) discuss such de-anonymization techniques, under the conditions of k-anonymity. Under

those conditions, each node is undistinguishable from other k-1 nodes, with regards to attributes which

may potentially identify an individual. This group of node constitutes a class of equivalence. While k-

anonymity can protect against identity disclosure, it fails to protect against attribute disclosure in the

general case. Another discussed notion of privacy is the l-diversity, which requires each class of

equivalence to contain at least l different values for a sensitive attribute. Authors propose instead t-

closeness as a measure of privacy, which requires that the distribution of an attribute in a class of

equivalence is close enough to that of the whole network, being less than a threshold t.

Weak passwords remain a major problem. Although, some service providers are either increasing

strictness of the conditions to accept a password as valid, experience shows that “lazy” users always

circumvent syntactical conditions (Yan et al., 2004). For example, in order to make passwords more

secure the providers started to require that at least one character should be a number. As a consequence,

many users simply appended a single digit to their weak password, without neither substantially

increasing the password security as a whole nor realising they are failing to do so.

Part of the problem is that, when choosing passwords, users are facing the choice between easy to

remember, easy to type passwords and secure passwords. As expected, users tend to choose the former.

Even if they are instructed with good techniques to build secure and memorable passwords (e.g., pass-

phrases), the large amount of services they subscribe creates the problem to remember which password

is needed for which account and, consequently, users tend to reuse their passwords over and over, with

terrible security implications. Also when longer pass-phrases are chosen, users prefer easier to

remember sequences, which make sense to them. But this motive also makes them vulnerable to more

advanced dictionary attacks, which gather phrases from popular books and online sources (Goodin,

2013).

Many solutions exist to improve password security, i.e., password managers and biometric

identification. However, these solutions have serious drawbacks. Biometric identification is

cumbersome because it may require specialised hardware and is often perceived as a privacy violation

on its own right. Additionally, biometric data, when used for authentication, are roughly equivalent to

strong and long passwords, which are hard to guess but which, on the other hand, are impossible to

change. This way, if a service requiring biometric data at login is broken, or behaves in a rogue way,

then the security based on those biometric data, on all sites, is made ineffective or completely broken.

Password managers, on the other hand, create a single point of failure that, if violated, compromises the

security of every single account the user had. In fact, this is not just a remote hypothesis (Slattery,

2011).

All considered, we think that one of the best strategies to improve the security of online social networks

would be using strong asymmetric cryptography to protect all the data stored inside the system.

Decrypting data protected with state-of-the-art cryptography is very hard. The system can be designed

so that data travels only in encrypted format and is decrypted only on the client machine, so that, in

essence no secret is revealed to a third party during regular usage. However, regular asymmetric

cryptography has high computational costs when messages have to be sent to a restricted audience that

varies its members during time, as in the case of a “group” or “circle” of friends (Canetti et al., 1999). The solution we favour is using flexible attribute-based encryption, such as CP-ABE (Bethencourt et

al., 2007). With attribute-based encryption, it is possible to create subordinate key-pairs associated with

arbitrary attributes. Such attributes can indicate belonging to some group or time-dependent access

rights. The system operators and other users have not necessarily access to the data so that, even if the

system is compromised, the data remains secure.

The main problem is that, without remotely storing the private key, it becomes difficult to access the

online social network from multiple devices, such as from home and from the office. In the specific

case of mobile-centric online social networks, it becomes possible to keep the private key on the

mobile device without impairing functionality. Backup and remote wipe procedures are widely

implemented and they can be used to solve most issues regarding losing the physical device.

Another positive effect of using flexible cryptography, such as attribute-based encryption, is that it

becomes possible to temporarily grant rights to another device from the main mobile device where the

private key is stored. In fact, this essentially allows using the mobile device as a non-intrusive hardware

key without actually having to copy the private key around.

The essential problem with encrypting data is that usually the system operators want to access such

data in order to improve target advertisement, that is their main source of revenue. Moreover,

information quality, flow experience and trust in OSNs users' loyalty are correlated; thus, system

operators have to pay close attention to them, to retain and increase their customers (Suki, 2012). So,

although technically feasible, an end-to-end encryption strategy would require either an entirely

different, although similarly profitable, business model or a less expensive architecture, typically

obtained moving from centralized to distributed.

There are two main categories of distributed social networks: federated and peer-to-peer (P2P). In a

federated system multiple entities cooperate to provide the service and each of them provides access to

the whole system to a subset of the total users. Each user can choose the federated provider that he

prefers, for example because he considers it worthier of trust. Service providers can also use high

security and privacy standards as a way to attract more users and, perhaps even publish the code open

source to increase confidence in their criteria. Direct access to the source code is possibly one of best

way for user to discover whether their provider is using inadequate security standards, and, in general

more competition and no lock-in are positive factors for increasing quality.

On the other hand, in a P2P system, every participant is both a user and a system provider at the same

time. However, purely P2P system can have issues with data availability, i.e., post from very badly

connected users can be difficult to obtain. Consequently P2P system may introduce “super-nodes” with

a role akin to that of a federated provider. The distinction between the two categories is in practice

blurred. The main advantage of P2P over federated approaches is that in the P2P scenario the code runs

on the user machine and, potentially, no information is transmitted outside with the explicit user

content. Moreover, trust management and negotiation mechanisms among users can be used in order to

improve the availability of the data even without introducing super-nodes (Tomaiuolo, Poggi &

Franchi, 2013).

Regardless of the actual form of distribution, the operative costs are essentially shared among multiple

entities, and, in the case of P2P the required resources are already provided by the users, so that most

costs are already covered.

Specifically in the field of social networking, various systems are being developed on the basis of peer-

to-peer communications and DHT indexing, including Safebook (Cutillo, Molva & Strufe, 2009),

PeerSoN (Bodriagov & Buchegger, 2013) and Blogracy (Franchi, Poggi & Tomaiuolo, 2013). Safebook

is based on a network of socially close peers, defined Matryoshka. Peers in a user's Matryoshka are

trusted and support the user by anonymizing communications and replicating content and profile

information. PeerSoN proposes a Broadcast Encryption protocol, where each recipient has a different key, which can be used to decrypt data received by the broadcaster . Blogracy, on the other hand, adopts

an Attribute-Based Encryption protocol for protecting access to users' content. It allows each user to

assign credentials to various groups of followers, for accessing protected content.

In conclusion, it should be noted that cryptography does not outright prevent “mass” phishing, however

it can make much more troublesome to discover enough data on users to conduct credible spear-

phishing attacks. Cloning phishing attacks are also less likely to succeed in federated networks, since

different federated providers would typically use different standards, and become meaningless in pure

P2P scenarios. Similarly, impersonation attacks cannot be directly avoided, but cryptography still

makes it harder to gather enough information to successfully conduct the attack. Another partial

defence against impersonation is that the same cryptographic keys that are used to guarantee

confidentiality can be used to guarantee authenticity, i.e., any message and any piece of information

can be guaranteed to come from the owner of the key: a clone could be easily detected.

Security questions are used in order to allow users to change their password in case it is stolen.

Depending on how the system is designed, there may not be a password to reset altogether. In these

cases, a similar problem is losing the private key. Although, in theory, backups should avoid the

problem, a failsafe mechanism may still be useful; however, because of the way the system works, it

should probably be something completely different, hopefully not vulnerable to the classic problems of

“security questions”. Moreover, in P2P system there is not even the concept of a central authority that

could “reset” any password or secret for the user and web of trust or similar strategies should be

employed.

Essentially, since all the pieces of information are encrypted, data published in online social networks

cannot be used to gather information to violate other systems. Moreover, even if the online social

network services were violated, the attackers would not access any data in clear.

CONCLUSION

In this work, we reviewed some typical social attacks that are conducted specifically on social

networking systems. In fact, a long with the large number of legitimate users, these systems are also

attracting the interest of crackers and scammers, who may seek information for triggering attacks to

even more sensible targets. Real-world examples of such violations are already available. Actually,

many factors concur to attract malicious actions: ( i ) the users’ carefree attitude in sharing information,

( ii ) the often inadequate security measures from the part of the system operators, and ( iii ) the high

value of the published information.

While the only lasting solution to privacy and security issues would be increasing the users' awareness,

much can and shall be done at the system level in order to protect the data with cryptography and to

decrease the impact of wrong choices and mistakes on the user’s part.

Security is a complex issue, especially considering that most online social network users are not willing

to be proactively engaged in the process. As a consequence, the system designer should pay additional

care so that most interactions have high security standards regardless of the user understanding of the

problem. However, for most online social networks providers’ security does not appear to be a priority,

possibly because it is a feature that does not help to “sell” the product. Unfortunately, every system that

is exploited makes every other system a bit less secure, because information leaks make information

attacks more likely to succeed. Moreover, security is going to be an increasingly important problem in

the future.

Therefore, we think that there is an expanding niche for smaller service providers willing to invest in

security and that they can actually enter the market using decentralised architectures in order to

decrease the costs of running the services. Such services could be designed from scratch in order to

provide good security standards, possibly using attribute-based encryption. REFERENCES

Acquisti, A., & Gross, R. (2009). Predicting Social Security numbers from public data. Proceedings of

the National academy of sciences , 106 (27), 10975-10980.

Agarwal, A. (2012). Security update & new features. Retrieved January 10, 2014, from

https://blog.dropbox.com/bre/07/security-update-new-features/

Bethencourt, J., Sahai, A., & Waters, B. (2007, May). Ciphertext-policy attribute-based encryption. In

Security and Privacy, 2007. SP'07. IEEE Symposium on (pp. 321-334). IEEE.

Bevand, M. (2010). Whitepixel. Retrieved January 10, 2014, from http://whitepixel.zorinaq.com/

Bilge, L., Strufe, T., Balzarotti, D., & Kirda, E. (2009, April). All your contacts are belong to us:

automated identity theft attacks on social networks. In Proceedings of the 18th international

conference on World wide web (pp. 551-560). ACM.

Bodriagov, O., & Buchegger, S. (2013). Encryption for peer-to-peer social networks. In Security and

Privacy in Social Networks (pp. 47-65). Springer New York.

Bonneau, J., Anderson, J., Anderson, R., & Stajano, F. (2009, March). Eight friends are enough: social

graph approximation via public listings. In Proceedings of the Second ACM EuroSys Workshop on

Social Network Systems (pp. 13-18). ACM.

Canetti, R., Garay, J., Itkis, G., Micciancio, D., Naor, M., & Pinkas, B. (1999, March). Multicast

security: A taxonomy and some efficient constructions. In INFOCOM'99. Eighteenth Annual Joint

Conference of the IEEE Computer and Communications Societies. Proceedings. IEEE (Vol. 2, pp. 708-

716). IEEE.

Cheng, J. (2012). Your iPhone calendar isn’t private—at least if you use the LinkedIn app . Retrieved

January 10, 2014, from http://arstechnica.com/apple/2012/06/your-iphone-calendar-isnt-privateat-least-

if-you-use-the-linkedin-app/

Cubrilovic, N. (2009). RockYou Hack: From Bad To Worse . Retrieved January 10, 2014, from

http://techcrunch.com/2009/12/14/rockyou-hack-security-myspace-facebook-passwords/

Cutillo, L. A., Molva, R., & Strufe, T. (2009). Safebook: A privacy-preserving online social network

leveraging on real-life trust. Communications Magazine, IEEE , 47 (12), 94-101.

Dai, W. (2009). Crypto++ 5.6.0 Benchmarks . Retrieved January 10, 2014, from

http://www.cryptopp.com/benchmarks-amd64.html

Dragusin, R. (2012). Data breach at IEEE.org: 100k plaintext passwords. Retrieved January 10, 2014,

from http://ieeelog.com/

Felt, A., & Evans, D. (2008). Privacy protection for social networking APIs. 2008 Web 2.0 Security and

Privacy (W2SP’08) .

Franchi, E., Poggi, A., & Tomaiuolo, M. (2013). Open social networking for online collaboration.

International Journal of e-Collaboration (IJeC) , 9 (3), 50-68.

Goodin, D. (2013). How the Bible and YouTube are fueling the next frontier of password cracking .

Retrieved January 10, 2014, from http://arstechnica.com/security/2013/10/how-the-bible-and-youtube-

are-fueling-the-next-frontier-of-password-cracking/

Honan, M. (2012). How Apple and Amazon Security Flaws Led to My Epic Hacking. Retrieved January 10, 2014, from http://www.wired.com/gadgetlab/2012/08/apple-amazon-mat-honan-hacking

Irani, D., Webb, S., Pu, C., & Li, K. (2011). Modeling unintended personal-information leakage from

multiple online social networks. Internet Computing, IEEE , 15 (3), 13-19.

Li, N., Li, T., & Venkatasubramanian, S. (2007, April). t-closeness: Privacy beyond k-anonymity and l-

diversity. In Data Engineering, 2007. ICDE 2007. IEEE 23rd International Conference on (pp. 106-

115). IEEE.

Merrill, M., & Beck, B. (2012). League of Legends Account Security Alert. Retrieved January 10, 2014,

from http://euw.leagueoflegends.com/news/league-legends-account-security-alert

Morris, R., & Thompson, K. (1979). Password security: A case history. Communications of the ACM ,

22 (11), 594-597.

Oechslin, P. (2003). Making a faster cryptanalytic time-memory trade-off. In Advances in Cryptology-

CRYPTO 2003 (pp. 617-630). Springer Berlin Heidelberg.

Provos, N., & Mazieres, D. (1999, June). A Future-Adaptable Password Scheme. In USENIX Annual

Technical Conference, FREENIX Track (pp. 81-91).

Rabkin, A. (2008, July). Personal knowledge questions for fallback authentication: Security questions

in the era of Facebook. In Proceedings of the 4th symposium on Usable privacy and security (pp. 13-

23). ACM.

Robertson, A. (2012). LulzSec hackers post data on 8,000 Twitter accounts, but your passwords are

safe . Retrieved January 10, 2014, from http://www.theverge.com/2012/6/12/3080534/lulzsec-reborn-

twitter-tweetgif-hack

Slattery, B. (2011). LastPass, Online Password Manager, May Have Been Hacked. Retrieved January

10, 2014, from

http://www.pcworld.com/article/227223/LastPass_Online_Password_Manager_May_Have_Been_Hack

ed.html

Strater, K., & Lipford, H. R. (2008, September). Strategies and struggles with privacy in an online

social networking community. In Proceedings of the 22nd British HCI Group Annual Conference on

People and Computers: Culture, Creativity, Interaction-Volume 1 (pp. 111-119). British Computer

Society.

Stroud, D. (2008). Social networking: An age-neutral commodity—Social networking becomes a

mature web application. Journal of Direct, Data and Digital Marketing Practice , 9 (3), 278-292.

Suki, N. M. (2012). Correlations of Perceived Flow, Perceived System Quality, Perceived Information

Quality, and Perceived User Trust on Mobile Social Networking Service (SNS) Users’ Loyalty. Journal

of Information Technology Research (JITR) , 5 (2), 1-14.

Tomaiuolo, M., Poggi, A., & Franchi, E. (2013). Supporting Social Networks With Agent-Based

Services. International Journal of Virtual Communities and Social Networking (IJVCSN) , 5 (1), 62-74.

Vijayan, J. (2007). TJX data breach: At 45.6M card numbers, it's the biggest ever. Retrieved January

10, 2014, from http://www.computerworld.com/s/article/9014782

Yan, J., Blackwell, A., Anderson, R., & Grant, A. (2004). Password memorability and security:

Empirical results. Security & Privacy, IEEE , 2 (5), 25-31.

Yin, S. (2012). Last.FM Joins eHarmony, LinkedIn to Celebrate Breach Week . Retrieved January 10, 2014, from http://www.pcmag.com/article.aspx/curl/2405492

Zheleva, E., & Getoor, L. (2009, April). To join or not to join: the illusion of privacy in social networks

with mixed public and private user profiles. In Proceedings of the 18th international conference on

World wide web (pp. 531-540). ACM.

Zhou, B., Pei, J., & Luk, W. (2008). A brief survey on anonymization techniques for privacy preserving

publishing of social network data. ACM SIGKDD Explorations Newsletter , 10 (2), 12-22.View publication statsView publication stats