1. Watermarking Relational Databases CSC 574/474 Information System Security

2. Cryptography Vs. Steganography Cryptography Encryption: translate information into an unintelligible form Decryption: decode to retrieve information Attackers cannot recover the information Stenography Hide information in a seemingly common message “Security through obscurity”: Attackers don’t know where to find the information

3. Steganography Examples Greek messengers Message tattooed into shaved head Invisible ink in a cover letter Bits hidden in pictures Sounds familiar? Hide one image into another Least significant bits Other forms?

4. Example Taken from http://www.petitcolas.net/fabien/steganography/image%5Fdowngrading/old/

5. Example Courtesy: http://www.petitcolas.net/fabien/steganography/image%5Fdowngrading/old/

6. Example Courtesy: http://www.petitcolas.net/fabien/steganography/image_downgrading/index.html

7. Example Courtesy: http://www.petitcolas.net/fabien/steganography/image_downgrading/index.html

8. Illustration of A Steganographic System http://www.vu.union.edu/~shoemakc/watermarking/watermarking.html

9. Digital Watermarks Insert marks into original data Use to demonstrate ownership: images, video, audio, software… Other usage? Should not significantly affect quality of original data Should not be able to be destroyed easily Deter instead of prevent illegal copying

10. Watermarking Databases Why? Data in database are intellectual properties Is it possible? Some numerical data do not need to be precise to be useful Example? Some data are imprecise in nature Example?

11. What Makes Watermarking Databases Different Dealing with multiple objects (tuples) instead of one Tuple order does not matter After dropping part of the database, the remaining part is still valuable

12. Desirable Features Detectability Allow undetectable marks Robustness Benign updates, malicious attacks Incremental updatability Do not need to re-compute watermarks during updates

13. Desirable Features Imperceptibility Preserve usefulness of the database Blind system Do not need the original database for detection Key-based system Watermarking scheme is open Only the private key matters

14. Attacks Benign updates Malicious attacks Bit attack Rounding attack Subset attack Mix and match attack Additive attack Invertibility attack

15. Basic Setup n tuples, v numerical attributes, P: primary key e least significant bits 1/r: fraction of tuples marked w: number of marked tuples (n/r) a: confidence parameter t: min number of correct mars for detection H: a one way hash function, K: private key, F: a MAC function F(m) = H(K || H(K||m))

16. Watermark Insertion Algo. PA1A2… Av Ai… A2A1P (1) if (F(P) mod r = 0) then should mark (2) choose to mark Ai where i = F(P) mod v bjbj bkbk b k-1 …b e-1 ……b1b1 (3) choose jth bit to mark where j = F(P) mod e b j = 0 if H(K || P) is even b i =1 otherwise

17. Watermark Detection Algo. 1. Determine whether a tuple is marked 2. Determine which attribute is marked 1. totalcount++ 3. Determine which bit is marked 4. Check whether the jth bit is the same as the expected mark 1. Matchout++ 5. Check whether a threshold t is met How to determine threshold t?

18. Operations on Watermarked Databases Query ? Updates? Insertion Deletion Modification

19. How to Determine Threshold t 1. The probability that b j is not changed by watermarking is __ 2. Out of w checks, the probability that t matches by chance is __ 3. What is the probability the detection algorithm makes a wrong decision? b j = 0 if H(K || P) is even b i =1 otherwise

20. How to Determine Threshold t 1. 0.5 2. C(w, t) * 0.5^w 3. (C(w, t) + C(w,t+1) + … + C(w,w)) * 0.5^w (1) Let a be the tolerable error rate, we have to choose the minimum t such that (1) < a

21. Robustness Against Attacks Bit-Flipping attack Choose s tuples from n tuples, flip all the e least significant bits, the chance to erase the watermark is Sum i=w-t+1,…,w C(w, i)C(n-w, s-i)/C(n,s)

22. Mix-and-Match Attack Mallory takes k fraction of the database Mix it with his own relation Create a new relation of size n For Alice to detect the watermark K*n/r + 0.5*(1-k)*n/r >= t

23. Additive Attack Mallory inserts his own watermark in Alice’s database How to determine who is the original owner? If two watermarking scheme marks the same bit of the same tuple Then?

24. Invertibility Attack Mallory finds a key that yields a satisfactory watermark on the database Affected by a The larger a is, is it easier or harder to find such a key?

25. Design Tradeoffs ↓ a↓ false hits↑ missed watermark ↓ r↑ robustness↑ data errors ↑ v↑ robustness ↑ e↑ robustness↑ data errors

26. Comments of the Paper Simple yet effective idea Thorough analysis Coming up with a good approach is hard Analyze, validate and make the approach complete is even harder No data on key length and hash function. What are their impact on performance?

27. Discussion Possible attacks Frequent updates of the same tuple? Side channels Water marking a tuple requires extra time Basic assumption The owner’s database is secured Regulations or law regarding database copyright?

28. Discussion How to handle non-numerical data Every change is significant But we have to make changes Minimize number of changes Encode message in cross-tuple properties E.g., attribute frequency histogram

29. Discussion Watermarking semi-structured data, e.g., XML? Attributes or element values can be similarly watermarked Define key is an issue The structure of the semi-structured data may also need to be watermarked

30. Further Reading Watermarking Relational Databases by Rakesh Agrawal and Jerry Kiernan, International Conference on Very Large Data Bases (VLDB), 2002. Rights Assessment for Discrete Digital Data, Ph.D thesis, by Radu Sion, Purdue University.