Evasion Attack (ML).

An inference-time attack in which an adversary crafts inputs that bypass a deployed machine-learning model's intended decision, such as evading a malware classifier or content filter. It belongs to the AI & ML Security category of cybersecurity.

An inference-time attack in which an adversary crafts inputs that bypass a deployed machine-learning model's intended decision, such as evading a malware classifier or content filter.

Evasion attacks operate after a model is trained and deployed: the attacker does not touch the training pipeline but manipulates queries to slip past detection. Most use adversarial examples, but the family also includes simpler bypass tactics like polymorphic malware, character obfuscation against text moderation, voice-cloning against speaker verification, or transformations against perceptual hashing. The NIST AI 100-2 report categorizes evasion as one of the four main adversarial ML threat classes alongside poisoning, privacy, and abuse. Defences include adversarial training, robust feature engineering, ensemble or multi-modal detection, runtime input sanitization, telemetry on confidence drift, and tight access controls on model APIs to limit query-based reconnaissance.

Defences for Evasion Attack (ML) typically combine technical controls and operational practices, as detailed in the full definition above.

Common alternative names include: Inference-time attack, Model evasion.