Adversarial Examples for Proof of Learning
Published:
An implementation of the spoofing attack that Proof-of-Learning has to survive, kept as a working reference rather than as a result.
The premise is short. Proof-of-Learning verifies that a model was trained by replaying part of the recorded trajectory, so an attacker who holds only the final weights has to manufacture a path that could plausibly have produced them. Adversarial-example techniques are one way to try, and having the attack in runnable form is what makes a defence testable rather than asserted.
My own work on the defence side is published: Enhancing Security of Proof-of-Learning Against Spoofing Attacks Using Feature-Based Model Watermarking and SecurePoL, with the dissertation at Enhancing Proof-of-Learning Security Against Spoofing Attacks Using Model Watermarking. The Proof-of-Learning lab film is the two-minute version of the argument.