Engineering / research
Test models against adaptive perturbations
Define attacker capabilities and success conditions before measuring perturbation attacks, defenses, transfer, and the scope of robustness certificates.
Your route through engineering
Follow the path or begin with a question. Every lesson is open.
Your place, at your pace. Progress is optional and saved only in this browser when available.
Read / explore / reason
Inside this module
Module 01 research
Test models against adaptive perturbations
Define attacker capabilities and success conditions before measuring perturbation attacks, defenses, transfer, and the scope of robustness certificates.
- 01
State exactly which perturbations count
Specify attacker access, coordinate units, feasible inputs, and success before interpreting any robustness score.
28 min - 02
Turn local derivatives into feasible test candidates
Derive one-step norm-constrained loss ascent and validate the candidate through the complete prediction path.
28 min - 03
Search the feasible region without drifting out of it
Implement projected iterative loss ascent with original-input constraints, best-candidate tracking, and reproducible restarts.
28 min - 04
Evaluate the defense that is actually deployed
Include preprocessing, rejection, randomness, and state in the attacked computation and validate candidates through its real forward path.
28 min - 05
Investigate gradients that conceal accessible failures
Use controlled diagnostics to distinguish genuine resistance from uninformative derivatives, saturation, and attack implementation errors.
28 min - 06
Measure robustness under a real information budget
Separate surrogate transfer, score-based probing, and label-only access while accounting for every target query.
28 min - 07
Train against a declared perturbation problem
Distinguish approximate robust optimization, explicit clean-adversarial mixtures, and distribution-consistency penalties.
28 min - 08
Certify only the classifier and noise law you proved
Connect Gaussian smoothing, conservative probability bounds, abstention, and finite-sample confidence to a scoped Euclidean certificate.
28 min - 09
Check that the perturbation preserves the task
Define transformation families and label adjudication without equating pixel distance, human similarity, or model agreement.
28 min - 10
Release a robustness result that others can audit
Bundle threat definitions, immutable artifacts, diverse attack evidence, and certificate scope into a reproducible test release.
28 min