A Bag of Tricks for Defending Deep Learning-based Face Recognition Pipelines against Backdoor Attacks

Summary

This technical report addresses backdoor attacks on multi-stage Face Recognition Systems, where training data comes from possibly-compromised third parties while training stays in-house. The authors present three defense techniques against data poisoning-based backdoors: one detects poisoned identities during training or before deployment, and two filter suspicious inputs at inference time. Each is described with its motivation, algorithms and computational impact, and evaluated on realistic face recognition workloads.

Mitigation

The report presents three defense techniques: one training-time method that detects poisoned identities during training or prior to deployment, and two inference-time methods that filter suspicious inputs before an answer is returned.

Topics

• Adversarial machine learning

Related items

• MediumResearchers Uncover Prompt Adware Hidden in OpenAI Codex Jailbreak ToolkitSimilar attack · The Hacker News
• InfoTop agent skill security resources — November 2026Similar attack · Adversa AI Blog
• HighCVE-2026-118627: MLflow SHAP load_explainer untrusted deserialization enables code executionSimilar attack · NVD/CVE Database
• HighCVE-2026-118623: MLflow PickleEvaluationArtifact remote code execution via untrusted pickleSimilar attack · NVD/CVE Database
• HighCVE-2026-88968: Claude Code Edit tool symlink race overwrites files outside workspace on macOSSimilar attack · NVD/CVE Database