Adversarial Robustness
Reverse engineering attacks through sparse recovery and generative-model inversion.
Reliable learning systems need methods for understanding and recovering from adversarial perturbations. My work studies how structured representations and generative priors can help separate a clean signal from an attack and identify the underlying threat model.
Reverse engineering attacks with sparse models
Our ICML 2022 work formulates reverse engineering of adversarial attacks as a block-sparse recovery problem. Under geometric conditions on the signal and attack subspaces, the method can recover the clean signal and attack, classify the signal, and identify the attack type.
Generative priors for nonlinear data
Our GAN inversion work extends this perspective to clean data modelled by a generative network. It combines latent inversion with block-sparse recovery and establishes deterministic linear convergence guarantees under the paper’s conditions. This connects adversarial robustness with the optimization geometry of generative models.
Selected publications
- Darshan Thaker, Paris Giampouras, and René Vidal. Reverse Engineering $\ell_p$ Attacks: A Block-Sparse Optimization Approach with Recovery Guarantees. ICML 2022, equal contribution by Darshan Thaker and Paris Giampouras.
- Darshan Thaker, Paris Giampouras, and René Vidal. A Linearly Convergent GAN Inversion-based Algorithm for Reverse Engineering of Deceptions. arXiv, 2023.