The field of machine learning is moving towards developing more robust and secure models, with a focus on adversarial robustness and defense. Recent papers have introduced new methods for defending against adversarial attacks, such as DRIFT, which uses a stochastic ensemble of lightweight filters to disrupt gradient consensus, and MANI-Pure, which uses magnitude-adaptive noise injection to suppress adversarial perturbations. Other papers have explored the use of diffusion models for secure and reversible face anonymization, and the development of new attack methods, such as DIA, which targets the integrated DDIM trajectory path. Noteworthy papers include DRIFT, which achieves substantial robustness gains on ImageNet, and VAGUEGAN, which introduces a stealthy poisoning and backdoor attack pipeline on image generative pipelines. Overall, the field is moving towards developing more robust and secure models, with a focus on adversarial robustness and defense.
Advances in Adversarial Robustness and Defense
Sources
Scalable Second-order Riemannian Optimization for $K$-means Clustering
Decoding Deception: Understanding Automatic Speech Recognition Vulnerabilities in Evasion and Poisoning Attacks
Pushing Toward the Simplex Vertices: A Simple Remedy for Code Collapse in Smoothed Vector Quantization
DRIFT: Divergent Response in Filtered Transformations for Robust Adversarial Defense
VAGUEGAN: Stealthy Poisoning and Backdoor Attacks on Image Generative Pipelines
Score-based Membership Inference on Diffusion Models
MANI-Pure: Magnitude-Adaptive Noise Injection for Adversarial Purification
Field Calibration of Hyperspectral Cameras for Terrain Inference
OmniDFA: A Unified Framework for Open Set Synthesis Image Detection and Few-Shot Attribution
PUREVQ-GAN: Defending Data Poisoning Attacks through Vector-Quantized Bottlenecks
Reconcile Certified Robustness and Accuracy for DNN-based Smoothed Majority Vote Classifier
Indirect Attention: Turning Context Misalignment into a Feature
Nearest matrix with multiple eigenvalues by Riemannian optimization
DiVeQ: Differentiable Vector Quantization Using the Reparameterization Trick
Stealing AI Model Weights Through Covert Communication Channels
CODED-SMOOTHING: Coding Theory Helps Generalization
Understanding Sensitivity of Differential Attention through the Lens of Adversarial Robustness
DIA: The Adversarial Exposure of Deterministic Inversion in Diffusion Models
Towards Adversarial Training under Hyperspectral Images
Secure and reversible face anonymization with diffusion models
HVAC-EAR: Eavesdropping Human Speech Using HVAC Systems
EditTrack: Detecting and Attributing AI-assisted Image Editing
Purrception: Variational Flow Matching for Vector-Quantized Image Generation
Towards Imperceptible Adversarial Defense: A Gradient-Driven Shield against Facial Manipulations
Mirage Fools the Ear, Mute Hides the Truth: Precise Targeted Adversarial Attacks on Polyphonic Sound Event Detection Systems