The field of deep learning is moving towards improving adversarial robustness and efficient model updates. Recent developments have focused on designing models that can withstand adversarial attacks, which are small perturbations to the input data that can cause the model to misbehave. Researchers have proposed various methods to achieve this, including risk-calibrated approaches to streaming intrusion detection and novel Wasserstein distributional attacks. Additionally, there is a growing interest in efficient model updates, such as transferring knowledge across pre-trained models and updating models using only a handful of labeled samples. Noteworthy papers in this area include 'Risk-Calibrated Bayesian Streaming Intrusion Detection with SRE-Aligned Decisions' and 'Gradient-Sign Masking for Task Vector Transport Across Pre-Trained Models', which demonstrate improved precision-recall and significant performance gains on vision and language benchmarks, respectively.
Advances in Adversarial Robustness and Efficient Model Updates
Sources
Risk-Calibrated Bayesian Streaming Intrusion Detection with SRE-Aligned Decisions
Gradient-Sign Masking for Task Vector Transport Across Pre-Trained Models
Tight Robustness Certificates and Wasserstein Distributional Attacks for Deep Neural Networks
The Easy Path to Robustness: Coreset Selection using Sample Hardness
Exploring and Leveraging Class Vectors for Classifier Editing
Adversarial Attacks Leverage Interference Between Features in Superposition
DRL: Discriminative Representation Learning with Parallel Adapters for Class Incremental Learning
A Function Centric Perspective On Flat and Sharp Minima
Topological Signatures of ReLU Neural Network Activation Patterns
KoALA: KL-L0 Adversarial Detector via Label Agreement
Sample-Centric Multi-Task Learning for Detection and Segmentation of Industrial Surface Defects
Injection, Attack and Erasure: Revocable Backdoor Attacks via Machine Unlearning
Generalist++: A Meta-learning Framework for Mitigating Trade-off in Adversarial Training
Towards Adversarial Robustness and Uncertainty Quantification in DINOv2-based Few-Shot Anomaly Detection
Deep Edge Filter: Return of the Human-Crafted Layer in Deep Learning
An Information Asymmetry Game for Trigger-based DNN Model Watermarking
When Flatness Does (Not) Guarantee Adversarial Robustness
TED++: Submanifold-Aware Backdoor Detection via Layerwise Tubular-Neighbourhood Screening
Structured Universal Adversarial Attacks on Object Detection for Video Sequences
Backdoor Unlearning by Linear Task Decomposition