N-Secure Fingerprinting for Copyright Protection of Multimedia 2004. 08. 23 Won-gyum Kim
Contents Watermarking vs. Fingerprinting Collusion attacks Collusion-secure fingerprinting code N-secure fingerprinting code Experimental results Conclusion
Watermarking vs. Fingerprinting Information hiding technique to protect copyright protection of multimedia Watermarking Embed owner’s information Protect owner’s copyright Only one watermarked content Fingerprinting Embed customer’s information Trace customer who re-distributes contents illegally Many different fingerprinted contents
Watermarking vs. Fingerprinting Distribute same contents O Customer 1 Content O Customer 2 . . . Owner’s information O Customer N
Watermarking vs. Fingerprinting Distribute different contents 1 Customer 1 Content 2 Customer 2 . . . Customer’s information N Customer N
Collusion attacks Use differences among fingerprinted contents Averaging attack Min-Max attack Negative correlation attack Zero-correlation attack
Collusion Attack Averaging Attack Min-Max Attack Average fingerprinted contents together Min-Max Attack Average min and max value of the fingerprinted contents Embedding and retrieving in the frequency domain are similar to that in the spatial domain, but it is different in the point that the watermark is embedded in the frequency domain after FFT. At first, transform the original image into the frequency domain using FFT Next, multiply bit sequence generated by secret key with seal image and embed the results to amplitude value Then recover original image using inverse FFT. Retrieving process is also similar to that in the embedding process. Firstly, transform watermarked image into the frequency domain using FFT and estimate approximate original from amplitude value. Then finally, extract seal image by multiplying the difference value with bit sequence generated by secret key.
Averaging attack ………… N Averaging image 1 2 As a second experiment, above is the result of embedding and extracting seal image in the frequency domain.
Collusion Attack Negative-correlation attack Zero-correlation attack Use median value Zero-correlation attack Use a target fingerprinted content to compare
Collusion-secure FC Marking Assumption The aim of collusion-secure FC By colluding, users can detect a specific mark if it differs between their copies; otherwise a mark can not be detected. The aim of collusion-secure FC After colluding, identify all colluders or at least more than one colluder
Collusion-secure FC To make robust to collusion Basic idea After colluding, the location of detectible code is unique according to all combinations of collusion Basic idea For 3 customers and 2 colluders C1 : 1 0 1 Collude C1 & C2 : 1 0 0 C2 : 1 1 0 Collude C2 & C3 : 0 1 0 C3 : 0 1 1 Collude C1 & C3 : 0 0 1
Collusion-secure FC 2-detecting code (Dittman, 2000) Based on the finite projective space Code for 3 customers with 2 colluders C1 : 1 0 0 0 1 1 0 C2 : 1 1 1 0 0 0 0 C3 : 0 0 1 1 1 0 0 Fingerprint 1 Fingerprint 2 Fingerprint 3 Points Lines 1 2 6 7 5 4 3
N-secure FC Use the location of undetectable code Content ID + Customer ID Code length is N+1 Content ID Customer ID
N-secure FC Code example For 7 customers with 7 colluders
All customers : 1 X X X X X X X N-secure FC Collusions X : undetectable C1 & C2 : 1 X X 1 1 1 1 1 C2 & C5 & C7 : 1 1 X 1 1 X 1 X C3 & C4 & C5 & C6 : 1 1 1 X X X X 1 All customers : 1 X X X X X X X
Embedding HVS MF(α) Original Image Fingerprinted Image W Code Customer index Code Generator Shuffle Key Produce pattern
Extracting Fingerprinted Image Correlator Filter De-shuffle Key Customer index
Experimental Results 512x512 Gray-scale Lena image For 15 users with 15 colluders Use general watermarking scheme Embed fingerprinting code into spatial domain of the image Divide into blocks and use shuffling to improve security level Do not consider the other watermarking attacks (a) Customer 1 without collusion (b) Collude Customer 1 & 2 (c) Collude all customers (1…15)
Conclusion Proposed fingerprinting code robust to collusion attack of N customers Trace user(s), who joined collusion Content ID + Customer ID Code length is N+1 Further study Consider the other watermarking attacks JPEG, RST, Filtering... Reduce code length