In the blog, I summary the accepted papers from Usenix Security 2022 which are related to my research interests. Basically, Usenix Security 2022 has three accepted paper lists from summer, fall, and winter:
- Summer: https://www.usenix.org/conference/usenixsecurity22/summer-accepted-papers
- Fall: https://www.usenix.org/conference/usenixsecurity22/fall-accepted-papers
- Winter: https://www.usenix.org/conference/usenixsecurity22/winter-accepted-papers
Interesting Paper list
- Summer
- Back-Propagating System Dependency Impact for Attack Investigation
- ML-Doctor: Holistic Risk Assessment of Inference Attacks Against Machine Learning Models
- DeepDi: Learning a Relational Graph Convolutional Network Model on Instructions for Fast and Accurate Disassembly
- Distinguished Paper Award: Online Website Fingerprinting: Evaluating Website Fingerprinting Attacks on Tor in the Real World
- Rapid Prototyping for Microarchitectural Attacks
- On the Security Risks of AutoML
- Inference Attacks Against Graph Neural Networks
- WebGraph: Capturing Advertising and Tracking Information Flows for Robust Blocking
- Distinguished Paper Award: [Dos and Don’ts of Machine Learning in Computer Security](https://www.usenix.org/conference/usenixsecurity22/presentation/arp
- Label Inference Attacks Against Vertical Federated Learning
- Fall
- PatchCleanser: Certifiably Robust Defense against Adversarial Patches for Any Image Classifier
- Poisoning Attacks to Local Differential Privacy Protocols for Key-Value Data
- Transferring Adversarial Robustness Through Robust Representation Matching
- Distinguished Paper Award: Provably-Safe Multilingual Software Sandboxing using WebAssembly
- On the Necessity of Auditable Algorithmic Definitions for Machine Unlearning
- Mitigating Membership Inference Attacks by Self-Distillation Through a Novel Ensemble Architecture
- Membership Inference Attacks and Defenses in Neural Network Pruning
- How Machine Learning Is Solving the Binary Function Similarity Problem
- FLAME: Taming Backdoors in Federated Learning
- Winter
I will keeping summarizing the interesting papers here.
Poison Forensics: Traceback of Data Poisoning Attacks in Neural Networks
Paper Author: Shawn Shan, Arjun Nitin Bhagoji, Haitao Zheng, Ben Y. Zhao
Main Idea
The paper points out the impossibility for the defense of adaptive adversarial attacks and provides a traceback method to tell the adversarial examples.
Key insight
Experiments
Dos and Don’ts of Machine Learning in Computer Security
Paper Author: Daniel Arp, Erwin Quiring, Feargus Pendlebury, Alexander Warnecke, Fabio Pierazzi, Christian Wressnegger, Lorenzo Cavallarok, Konrad Rieck
Main Idea
This paper lists ten common issues in the AI security papers with the explanations and potential solutions. It also decorates them by analysizing the impact with examples.
The figure shows the ten pitfalls and their corresponding stages in the ML workflow.
Key insight
Pitfalls list
- Data collection and Labeling
- Sampling Bias: The collected data are not representitive enough to show the real data distribution.
- Label Inaccuracy: The ground truth labels are not accurate, stable.
- System Design and Learning
- Data Snooping: The training data are not available in practice.
- Spurious correlations: Unrelated shortcut patterns for separating classes.
- Biased Parameters: Some parameters are based on the test set instead of fixing at training time.
- Performance Evaluation
- Inappropriate Baselines: Without or with limited baseline methods is in appropriate.
- Inappropriate Measures: The performance measures do not account for the constraints of the application scenario.
- Base rate Fallacy: One class imbalance which does not consider when interpreting the performance measures.
- Deployment and Operation
- Lab-only Evaluation: The leaning model is evaluated in a lab environment only instead of discussing the practical limitations.
- Inappropriate Threat Model: The robustness of machine learning models is negelected such as poisoning and evasion attacks.
Impact Analysis
The paper works on four applications:
- Mobile malware detection
- Vulnerability discovery
- source code authorship attribution
- network intrusion detection
