SecurePoL with Watermarking
Published:
Official code and experiments for my PhD dissertation, Enhancing Proof-of-Learning Security Against Spoofing Attacks Using Model Watermarking.
Proof-of-Learning asks a model owner to prove they trained a model rather than downloaded it, by keeping the trajectory the training took. The attack it has to survive is spoofing: reconstructing a plausible trajectory backwards from a stolen set of final weights. The dissertation’s answer is a second lock, a watermark woven into the model itself, so that a verifier accepts only when the training path checks out and the mark is present. A forger can copy the curve; SecurePoL adds a watermark-consistency check, so the tested spoofing routes must satisfy both conditions.
The repository is a set of Jupyter notebooks covering the three watermarking strategies the work compares and the spoofing experiments they are tested against. The published versions are the IEEE Access paper and SecurePoL, and the watermarking lab film walks through the three strategies, what each needs before it can be verified, and the weakness they share.