Adversarial robustness needs a threat model
Research context
research method · established · As of 2026-10-10
A norm-bounded adversarial evaluation asks whether an attacker can change a model's prediction using perturbations within a specified set around an input.
Key knowledge
A norm-bounded adversarial evaluation asks whether an attacker can change a model's prediction using perturbations within a specified set around an input. Goodfellow and colleagues' gradient-sign method takes an L-infinity step based on the input loss gradient. Attack success is empirical evidence for vulnerability under that model; failure of a finite attack search is not proof of robustness. Norm, input scaling, attacker knowledge and physical/perceptual relevance determine the claim.
Values and equations
x_adv=x+epsilon sign(grad_x ell(theta,x,y)), with valid-input projection when required
- x
- Original input
- epsilon
- L-infinity perturbation budget in input units
- ell
- Differentiable attack loss
- theta
- Model parameters
Conditions and scope
- Specify norm/radius, valid-input projection, targeted versus untargeted goal, white/black-box access and attack budget.
- Adversarial perturbations, distribution shift and random corruption are different evaluation populations.
Sources · 1
References
Sources retain their recorded scope. A source link alone does not establish verification of every statement.
Evidence and review
Author checked this English statement, displayed equation where present and stated hypotheses against these source locations: §§2–5 norm-bounded perturbations, linear explanation and gradient-sign method. The check is bounded to the recorded claims; no complete proof reconstruction, software proof replay or universal literature completeness is claimed.
2026-10-10 · Author source-body/statement/equation/domain check, with graph associations scoped individually. Separate independent content/source/inventory review is required before branch closure.
Entry details
Permanent link to this entry ↗- Record identity
- research-info-adversarial-threat-model
- Revision
- research.1
- Structured reference record
- Postgraduate & research reference