# Analyzing Federated Learning through an Adversarial Lens

**Authors:** Arjun Nitin Bhagoji, Supriyo Chakraborty, Prateek Mittal, Seraphin Calo  
**arXiv:** [1811.12470](https://arxiv.org/abs/1811.12470)  
**Date:** 2018-11-29  
**Summary page:** https://www.summarizepaper.com/en/paper/1811.12470v4/

## Key points

- The paper explores federated learning and its vulnerability to adversarial attacks
- Federated learning distributes model training across multiple agents for privacy reasons
- Study focuses on model poisoning attacks by a malicious agent to cause misclassification
- Strategies investigated include boosting malicious agent updates and using minimization techniques
- Visual explanations of model decisions are indistinguishable between benign and malicious models
- Results emphasize the importance of robust defense strategies in federated learning

## Summary

The paper "Analyzing Federated Learning through an Adversarial Lens" delves into the realm of federated learning and its vulnerability to adversarial attacks. This method distributes model training across multiple agents to address privacy concerns, with each agent only sharing parameter updates for aggregation at a central server. The study focuses on exploring the impact of model poisoning attacks initiated by a single malicious agent, aiming to cause misclassification of specific inputs with high confidence. Various strategies are investigated, including boosting the update from the malicious agent and utilizing alternating minimization and parameter estimation techniques for attack stealth and success. Surprisingly, interpretability techniques reveal that visual explanations of model decisions are indistinguishable between benign and malicious models, highlighting the challenge in detecting adversarial behavior within federated learning settings. The results underscore the need for robust defense strategies to safeguard against such threats and shed light on vulnerabilities inherent in collaborative machine learning environments.