Intelligent E-Commerce System Lab. Aettie , Ji

1. ALPACAS: A Large-scale Privacy-aware Collaborative Anti-spam SystemZ. Zhong, L. Ramaswamy and K. Li, IEEE, INFOCOM 2008 Intelligent E-Commerce System Lab. Aettie, Ji

2. OUTLINE • INTORDUCTION • PRIOR WORK • THE ALPACAS ANTI-SPAM FRAMEWORK • Feature-Preserving Fingerprint • Privacy-Preserving Collaboration Protocol • System Structure • EXPERIMENTS & RESULTS • DISSCUSION • CONCLUSION

3. INRTODUCTION • Motivations • Recent spam attack expose strong challenges to statistical filters, which have been popular. • Collaborative spam filtering has a natural defense paradigm, wherein information of spam is shared, sincethe spammers sends similar emails to several target receivers. • However, privacy of participating collaboration is an important challenge. For protecting privacy, digest approaches have been proposed but they are not sufficient.

4. INRTODUCTION • Contributions • ALPACAS: Large-scale Privacy-Aware Collaborative Anti-spam System. • A resilient fingerprint generation technique, “feature-preserving transformation”, is proposed. • A privacy-preserving protocol is designed to control the amount of information to be shared. • The experimental results demonstrate that the ALPACAS outperforms traditional stand-alone statistical filters.

5. PRIOR WORK • Drawbacks of the existing collaborative anti-spam schemes (using DCC). • How it works? • Participating servers in DCC share the email’s digests computed through hash functions such as MD5. • DCC system replies back with the recent statistics about the digests. • Drawbacks • Hashing schemes like MD5 generate complete different hash value even if a single byte is altered. • The DCC scheme does not completely address the privacy issue.  inference-based privacy breaches.

6. THE ALPACAS ANTI-SPAM FRAMEWORK(1/2) • Challenges • To protect email privacy, • The messages have to be encrypted. • It should retain important feature of the messages. • To avoid inference-based privacy beaches, • It is necessary to minimize the information revealed during the collaboration. • ALPACAS framework components • Feature-preserving fingerprint • Privacy-preserving protocol • DHT-based architecture

7. THE ALPACAS ANTI-SPAM FRAMEWORK(2/2) (b) Internal mechanism of EA4 (a) ALPACAS Network Fig. 1: ALPACAS System Overview

8. THE ALPACAS ANTI-SPAM FRAMEWORK Feature-Preserving Fingerprint(1/4) • Shingle-based Message Transformation • Shingle: If two documents vary by a small amount their shingle sets also differ by a small amount. Fig. 2: ALPACAS Feature Sets, DCC and Razor Digests for 2 spam emails (Texts in bold font indicate differences)

9. THE ALPACAS ANTI-SPAM FRAMEWORK Feature-Preserving Fingerprint(2/4) • Shingle-based Message Transformation • Generation of transformed feature set of message Ma(TFSet(Ma)) • Computing Rabin fingerprint[11] of consecutive tokens in sliding window of length W • Each fingerprint is in the range of (0, 2K– 1) • For a message with X tokens, X – W + 1 fingerprints are obtained. • The smallest Y are retained. • The similarity between Ma and Mb can be calculated as

10. THE ALPACAS ANTI-SPAM FRAMEWORK Feature-Preserving Fingerprint(3/4) • Shingle-based Message Transformation • Inconsideration of the privacy preservation, • Rabin fingerprint algorithm is one-way hash function such that it is infeasible to reverse. • However, it is possible to infer a word or a group of words from an individual feature value.

11. THE ALPACAS ANTI-SPAM FRAMEWORK Feature-Preserving Fingerprint(4/4) • Term-level Privacy Preservation • Controlled shuffling • The email text is divided into consecutive h chucks of z consecutive token. • The tokens in each chuck are shuffled in a pre-defined manner, remaining the ordering of chucks. • Each chuck is divided into y sub-chuck. (y is a factor of z.) • The tokens in chuck CKh are shuffled such that the token at rth position in the sth sub-chuck is moved to (r ⅹ y + s)th position in CKh. • If two messages contain an identical term, by shuffling the term, the feature set could be different.

12. THE ALPACAS ANTI-SPAM FRAMEWORK Privacy-Preserving Collaboration Protocol (1/3) • Spam/ham dichotomy • Protocol • EAj receives Ma, then computes TFSet(Ma). • EAj sends query to other agent with subset of TFSet(Ma). • EAk receives the query, then check its spam/ham KB. • For each matching entry in spam KB, EAk sends back the complete transformed feature set. • For each matching entry in ham KB, EAk sends back a small, randomly selected part of the transformed feature set. • Revealing the contents of a spam email does not affect the privacy, whereas revealing information about a ham email constitutes a privacy breach.

13. THE ALPACAS ANTI-SPAM FRAMEWORK Privacy-Preserving Collaboration Protocol (2/3) Fig. 3: ALPACAS Protocol: Query and Response

14. THE ALPACAS ANTI-SPAM FRAMEWORK Privacy-Preserving Collaboration Protocol (3/3) • Protocol(cont’) • EAj now computes the ratio of MaxSpamOvlp(Ma) to MaxHamOvlp(Ma) and decides whether the Mais spam or ham. • If the score is greater than a threshold λ, Mais classified spam, otherwise ham.

15. THE ALPACAS ANTI-SPAM FRAMEWORK System Structure (1/2) • Design principle • DHT-based Architecture • EAj is responsible for maintaining information about all the emails whose TFSet as one feature element in the range of allocated to it. • A query should be sent to an email agent only if it has a reasonable chance of containing information about the email that is being verified. Contacting any other email agent not only introduces inefficiencies but also leads to unnecessary exposure of data.

16. THE ALPACAS ANTI-SPAM FRAMEWORK System Structure (2/2) • DHT-based Architecture (cont’) • N email agent. • All feature elements lie within (0, 2K-1). • The range (0, 2K-1) is divided into N overlapping region as {(MinF0,MaxF0), (MinF1,MaxF1), . . . , (MinFN-1, 2K−1)}. • (MinFj, MaxFj) denotes the sub-range allocated to EAj. • For spam, EAjstores the entire TFSet. • For ham, EAjstores the subset of TFSet. • If MinFj≤ Ft ≤ MaxFj, then EAj is called rendezvous agent of feature element Ft.

17. EXPERIMENTS & RESULTS • Benchmarked algorithm • Bogofilter based on Bayesian filtering • Calculating a spamminess score of the email. • DCC based on simple hash-based collaborative filtering • Counting the number of times the hash value of the email has been reported as a spam.

18. EXPERIMENTS & RESULTS Experimental Setup • Dataset • TREC email corpus & SpamAssassin email corpus • TREC corpus is classified into 67 email sets according to their target address (67 agents). • Half of each email set including ham and spam is used for training and the remainder for testing. • Each individual has a pre-classified email corpus(SpamAssassin) a the initial knowledgebase.

19. EXPERIMENTS & RESULTS Performance Metrics • Spam filtering accuracy • A ham email that is classified a spam by the filtering scheme is termed as false positive. • Privacy of collaborative anti-spam system • Message-level privacy breach percentage is defined as the ratio number of test ham messages suffering privacy compromises to the total number of test ham messages. • Communication overhead of the system • Per-test communication cost metric is defined as the total number of messages circulated in the system during the entire experiment.

20. EXPERIMENTS & RESULTS SPAM Filtering Effectiveness Fig. 6: System Overall Accuracy (DCC is not displayed because its FP is 0) Fig. 4: False Positive Percentages of ALPACAS, BogoFilter and DCC Fig. 5: False Negative Percentages of ALPACAS, BogoFilter and DCC

21. EXPERIMENTS & RESULTS Robustness Against Attacks Fig. 7: System Robustness Against Good-Word Attacks Fig. 8: System Robustness against Character Replacement Attacks

22. EXPERIMENTS & RESULTS Privacy Awareness Fig. 9: Privacy Breach in ALPACAS (Varying Number of Agents)

23. EXPERIMENTS & RESULTS Communication Oveheads Fig. 10: Communication Overheads of the ALPACAS and the DCC systems

24. EXPERIMENTS & RESULTS Massage Transformation Algorithm Analysis Fig. 11: False Positive of ALPACAS for Various Parameter Setup Fig. 12: False Negative of ALPACAS for Various Parameter Setup Fig. 13: Effectiveness of Controlled Shuffling Strategy

25. DISCUSSION • Approaches like statistical filtering combined the feature preservation transformation scheme. • Applying dynamic nature of email agent to the system using replication and finger-table based routing. • Approaches for preventing malicious email agents.

26. CONCLUSION • In this paper, the design and evaluation of ALPACAS is presented. • The two novel features: • A feature preserving transformation technique • A privacy-preserving protocol • Our initial experiments show that ALPACAS • Is very effective in filtering spam. • Has high resilience towards various attacks. • Has strong privacy protection to the participating entities.