#518 LACMUS Latent Concept Masking for General Robustness Enhancement of DNNs Hongsheng Hu
Security researchers examine technical aspects of #518 LACMUS Latent Concept Masking for General Robustness Enhancement of DNNs Hongsheng Hu.
Summary
Robustness problem
Presenters from UNSW, Monash, and CSIRO at IEEE S&P introduce LACMUS (latent concept masking) to improve deep neural network robustness against adversarial pixel perturbations, semantic shifts such as lighting changes, and out-of-distribution samples where standard training fails.
Limits of adversarial training
Classic adversarial training mixes benign and adversarial examples but tends to be attack-specific, hurts clean accuracy, relies on low-level noise hard to interpret, and needs many adversarial samples. They identify four research gaps driving their method.
LACMUS approach
LACMUS encodes inputs through a VQ-VAE-style encoder into vectors matched against a pretrained concept codebook, masks or replaces concepts, decodes perturbed samples, and treats misclassified reconstructions as conceptual adversarial examples for training. Experiments using under one percent of training data report maintained clean accuracy and broader resilience to pixel, semantic, and distribution-shift attacks on MNIST and related sets.
Related talks
IEEE Symposium on Security and Privacy·15 min
Security researchers examine technical aspects of High-Order Masking of Lattice Signatures in Quasilinear Time.
IEEE Symposium on Security and Privacy·16 min
Security researchers examine technical aspects of On the Robustness of Deep K-Nearest Neighbors.
IEEE Symposium on Security and Privacy·19 min
Presenters detail vulnerability discovery and exploitation in Asm2Vec: Boosting Static Representation Robustness for Binary Clone Search....