Interactive experiment

DFA, SIFA and observation channels

Choose a synthetic sample set and an observation. The control is a constant class comparison: each class repeats one constant value, and the two constants are compared. Control rejection and information leakage are separate questions. In this fixture, “valid” means the fault did not change the computation, so it survived the detector; “invalid” means the computation was affected and the detector rejected it. These are outcome-code labels, not a DUT output-valid pin. The A/B comparison below is only a generic teaching comparison of sample labels; it is not a SIFA attack or a key-recovery model. In SIFA, the attacker keeps only the ineffective faulted computations (a fault was induced but the output stays correct) and examines the conditional distribution, or bias, of an intermediate value within that subset; it does not compare effective and ineffective groups to recover a key. Effective or ineffective describes whether the computation or output changed under the stated oracle, not the DUT out_valid pin. The second part, a noisy sample exercise, uses a fixed seed and a held-out split to show how sample count and noise change a class comparison.

Controls

Sample set Observation Suppress faulty ciphertextSamples per class (synthetic exercise) Noise amplitude (± units) Held-out share of each class

Small synthetic samples for reasoning only; not an attack result, measured leakage, or security probability. The sample exercise is a deterministic teaching model (fixed seed 13), not measured leakage or an attack result.

Static worksheetNo-JavaScript comparison (constant class comparison): the separated set gives the two classes different constant codes, so this sample distinguishes them; the overlapping set gives both classes the same constant code, so this sample does not. For the DFA view the codes stand for correct and faulty ciphertext; for the outcome-label view, valid and invalid outcomes (a generic label comparison, not SIFA). In this fixture “valid” means the fault did not change the computation and survived the detector, and “invalid” means the computation was affected and the detector rejected it; neither is a DUT output-valid pin. A separation in this sample is not an attack result. This A/B comparison is only a generic teaching comparison of sample labels, not a SIFA attack or key-recovery model: SIFA conditions on ineffective faulted computations (fault induced, output still correct) and examines the bias of an intermediate value in that subset, and effective/ineffective describes computation or output change under the stated oracle, not DUT out_valid. The sample exercise (samples per class, noise, held-out share) needs JavaScript: it builds fixed-seed noisy samples, trains on the earlier ones, and scores the held-out ones. With zero noise the separated set classifies every held-out sample and the overlapping set stays at chance; more noise or fewer samples makes the held-out score less stable. It is a teaching model, not measured leakage or an attack result. Suppressing faulty data masks only the ciphertext channel; validity and timing remain separate observations.