Enhancing Model Robustness and Fairness in Adversarial Environments

The recent developments in the research area of machine learning and uncertainty quantification have shown a significant shift towards enhancing model robustness and fairness, particularly against adversarial attacks and uncertainty-based threats. There is a growing emphasis on integrating uncertainty quantification into model robustness guarantees, which allows for more systematic evaluations of network architectures and uncertainty measures. Additionally, the field is witnessing advancements in post-hoc calibration methods, with novel techniques like feature clipping being proposed to address overconfidence in deep neural networks. The use of mixed-precision quantization frameworks is also gaining traction, with a focus on not only maintaining accuracy but also ensuring certifiable robustness. Furthermore, there is a notable trend towards developing hardware-friendly data generation methods for post-training quantization, which is crucial for privacy and security constraints. The integration of uncertainty quantification and adversarial training is being explored to provide more secure and trustworthy uncertainty estimates, which is vital for security-sensitive applications. Overall, the field is moving towards more robust, fair, and reliable machine learning models that can handle complex and adversarial environments.

Noteworthy papers include one that introduces a novel mixed-precision quantization framework for certifiably robust DNNs, and another that proposes a post-hoc calibration method based on feature modification, significantly improving model calibration.

Sources

Interval-valued q-rung orthopair fuzzy Weber operator and its group decision-making application

Feature Clipping for Uncertainty Calibration

Classification under strategic adversary manipulation using pessimistic bilevel optimisation

Integrating uncertainty quantification into randomized smoothing based robustness guarantees

Inverting Gradient Attacks Naturally Makes Data Poisons: An Availability Attack on Neural Networks

On the Robustness of Adversarial Training Against Uncertainty Attacks

Data Generation for Hardware-Friendly Post-Training Quantization

Reliability Assessment of Information Sources Based on Random Permutation Set

Scoring Rules and Calibration for Imprecise Probabilities

FAIR-TAT: Improving Model Fairness Using Targeted Adversarial Training

Legitimate ground-truth-free metrics for deep uncertainty classification scoring

Wide Two-Layer Networks can Learn from Adversarial Perturbations

ARQ: A Mixed-Precision Quantization Framework for Accurate and Certifiably Robust DNNs

Built with on top of