Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Aug 10·5 min read
Bypassing Krum: Selection-Aware Backdoor Attacks in Federated Learning
TL;DR
Krum-Proxy is a novel backdoor attack method in federated learning that effectively bypasses robust aggregation techniques like Krum. By optimizing malicious updates to blend with benign data, it achieves higher attack success while maintaining clean model accuracy.
✦ Why It Matters
Evaluate your federated learning systems for vulnerabilities to selection-aware backdoor attacks today.
Key Takeaways
- Assess your federated learning models for susceptibility to Krum-Proxy attacks.
- Implement additional defenses against selection-aware adversaries.
- Consider alternative aggregation methods beyond distance-based techniques.
- Regularly update your security protocols to address emerging threats.
How It Works
Krum-Proxy operates by creating adversarial updates that closely resemble benign updates, allowing them to blend into the majority during aggregation. It employs a two-stage optimization process: first, it defines attack objectives specific to the task, and then it refines these updates geometrically to fit within the benign distribution.
Techniques like nearest-neighbor proxies and stochastic reference modeling help guide this refinement, ensuring that the malicious updates remain undetected.