NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·2h ago
AI Research
arXiv CS.AI·May 25·12 min read
Adversarial Vulnerability Under Temporal Concept Drift: A Longitudinal Study of Android Malware Detection
TL;DR
Machine learning models for detecting Android malware degrade over time as new malware evolves, but this robustness loss under temporal drift (data distribution changes) has never been systematically studied. Researchers evaluated adversarial robustness across 10+ years of Android apps using FGSM and SPSA attack methods under three realistic deployment scenarios, introducing metrics like RobustDrop and Adversarial Amplification Factor to measure robustness decay.
✦ Why It Matters
Engineers deploying malware detectors must account for temporal drift when assessing long-term robustness and plan retraining strategies accordingly.
Key Takeaways
- Evaluate your detection systems for adversarial vulnerabilities over time.
- Incorporate drift-aware metrics to assess long-term performance.
- Use expanding-window retraining to improve robustness against evolving threats.
- Monitor the impact of temporal gaps on detection accuracy regularly.
How It Works
The study employs a longitudinal approach, analyzing Android applications over a decade. It generates adversarial examples using techniques like Fast Gradient Sign Method (FGSM) and SPSA, assessing their impact on various classifiers.
New metrics, such as RobustDrop and Adversarial Amplification Factor (AAF), quantify the relationship between temporal drift and adversarial robustness.