V2: Model Security — MLASVS-MODEL¶

Overview¶

Model security addresses attacks that target the trained model itself. These include adversarial evasion (manipulating inputs to cause misclassification), extraction (stealing model parameters or behavior), inversion (reconstructing training data), and backdoor/trojan attacks (hidden malicious behaviors).

Key Threats¶

Subcategories¶

MODEL-1: Adversarial Robustness (Controls MODEL-001, MODEL-002, MODEL-003, MODEL-010, MODEL-012, MODEL-016, MODEL-017, MODEL-024, MODEL-025, MODEL-026, MODEL-027, MODEL-028, MODEL-029)¶

Testing and hardening against evasion attacks using perturbation-based inputs.

MODEL-2: Extraction Prevention (Controls MODEL-004, MODEL-005, MODEL-006, MODEL-018, MODEL-023)¶

Preventing adversaries from stealing model behavior or parameters.

MODEL-3: Inversion & Privacy (Controls MODEL-009, MODEL-011, MODEL-013, MODEL-019, MODEL-020)¶

Preventing adversaries from inferring training data from model outputs.

MODEL-4: Backdoor Detection (Controls MODEL-021, MODEL-022)¶

Detecting hidden malicious behaviors injected during training.

MODEL-5: Model Integrity (Controls MODEL-007, MODEL-008, MODEL-014, MODEL-015, MODEL-030)¶

Ensuring models are authentic, untampered, and properly versioned.

Control Inventory¶

L1 Controls (15)¶

L2 Controls (15)¶

Related EW Links¶

• MLASTG Test Cases: Model Security
• MLASWE-0001: Adversarial Perturbation
• MLASWE-0003: Model Extraction
• MLASWE-0004: Model Inversion
• MLASWE-0005: Membership Inference
• MLASWE-0007: Backdoor/Trojan