This post summarizes the state of the art on the privacy and security of machine learning from my naive understanding. Corrections are highly appreciated.

USENIX Security 2021

Tracks

• Machine Learning: Backdoor and Poisoning
• Machine Learning: Adversarial Examples and Model Extraction
• Adversarial Machine Learning: Defenses
• Machine Learning: Privacy Issues

Papers

• Systematic Evaluation of Privacy Risks of Machine Learning Models. Keywords: Bayes' theorem, class-dependent threshold.
• Defeating DNN-Based Traffic Analysis Systems in Real-Time With Blind Adversarial Perturbations. Keywords: constrained domain, DNN, network traffic, Tor.
• Extracting Training Data from Large Language Models. Keywords: privacy, GPT-2 (only public data, a safe attack), model extraction, sensitive data, model/neural network memorization, differential privacy.
• Explanation-Guided Backdoor Poisoning Attacks Against Malware Classifiers

Others

• CADE: Detecting and Explaining Concept Drift Samples for Security Applications. Keywords: concept drift, Android malware, distance-based explanation.
• Mind Your Weight(s): A Large-scale Study on Insufficient Machine Learning Model Protection in Mobile Apps. Keywords: model protection, mobile, apps.