Towards Randomized Signal Processing: The Randlet transform and Image Identification

Ramarathnam Venkatesan, Microsoft Research

We motivate and suggest the use of explicit randomization techniques in signal processing applications that involve attacks or unknown transformations that do not significantly alter the signal in a perceptual sense. This will be useful in search, identification and information embedding and watermarking applications. We introduce a signal transform that is chosen randomly (using a key) from a family, akin to universal hash functions in CS/cryptography.We discuss some of the issues of modelling and analysis. At an intuitive level the approach is to extract robust pseudo-random features of the signal that are that is hard to guess by an adversary without the secret key. We present algorithms and experimental results that show how such a signal representation can be used for perceptual hashing and for embedding information in signals using optimization algorithms.

Talk based on joint with M.Malkin (Stanford), and earlier works with M.Kivanc (Microsoft Research), S.Kozat (IBM Res), M.Kesal & T. Liu (Urbana Champaign)

Gates 4B (opposite 490) Friday 03/04/05 1630 hrs