Peer-to-peer (P2P) machine learning enables decentralized training, but is highly vulnerable to poisoning attacks, where compromised nodes inject malicious updates. Existing Byzantine-resilient rules, such as Krum or Trimmed Mean, rely on hard exclusion of suspected updates, which reduces diversity and may discard benign but atypical contributions. We propose a soft-weighting defense for P2P learning that adaptively down-weights suspicious updates using an anomaly detection model trained on statistical and temporal indicators. This design mitigates the influence of malicious contributions while preserving diversity. Preliminary results show that our approach improves robustness against untargeted poisoning over classical aggregation baselines while maintaining scalability and adaptability in decentralized settings.