Side channels: define what an adversary can observe

Let a secret variable describe the information the defender intends to keep private, and define exactly what the other tenant can observe: its own operation latency, cache-related timing, shared-service response time, or another permitted signal. Specify whether the adversary can run native code, choose placement, invoke a cache-flush instruction, obtain a fine-grained clock, or repeat an experiment. The existence of a shared resource alone is not a complete attack. The observation must carry relevant secret-dependent information under a feasible and authorized test model.

Original Flush+Reload research uses shared memory pages and cache-line flushing to observe victim activity under its stated conditions. Prime+Probe-style models instead study contention in cache sets and need not require the same shared virtual page. Do not generalize one technique’s prerequisite into a universal requirement for all cache channels. Different processors, scheduling, virtualization, timers, page sharing, and mitigation settings change the feasible observation. A defensive assessment records these conditions so that a demonstrated channel and a failed reproduction can both be interpreted precisely.

In an authorized fixture, vary secret labels while controlling unrelated work, and reserve blind evaluation traces distinct from tuning data. Otherwise a classifier can memorize run identifiers, request sizes, machine temperature, or the order in which fixtures were executed. Balanced labels make a simple baseline visible, but they do not remove nuisance variables. Pair or randomize experiments, preserve the data-generation record, and check whether the proposed mitigation changes the secret-dependent mechanism rather than merely changing a convenient incidental marker.

Report false positives, false negatives, class balance, sample size, and uncertainty, not just one accuracy percentage. A classifier exceeding chance on a holdout fixture supplies evidence for that fixture and observation model. It does not automatically establish key recovery, arbitrary victim inference, or a production exploit. Conversely, chance-level results may reflect too little data, a weak observer, an irrelevant feature, or a channel outside the chosen observation. Match the conclusion to the tested capability instead of treating one statistical result as a complete confidentiality theorem.

Possible defenses include removing a particular sharing condition, using secret-independent implementations, partitioning an implicated resource, or changing placement and trust assignments. Each has a bounded claim. Disabling shared pages addresses a prerequisite of a specific Flush+Reload setup but does not remove every contention channel. Cache-way allocation is not a proof against all memory, timing, power, or shared-service observations. Define what changes, what remains shared, and which adversary capabilities are excluded, then repeat the evaluation alongside performance and correctness tests.

Reset the probability lab, explicitly set p=0.5 and n=20, predict the mean number of correct guesses, then Play or Step. For a separate hypothetical observer Reset and explicitly set p=0.8 and n=20 before predicting again. The means are ten and sixteen under the assumed independent model. Playback highlights exact binomial outcomes; it neither draws attack traces nor estimates p from observed accuracy. The lab therefore illustrates a stated probabilistic hypothesis, while the actual defensive experiment must justify labels, independence, observation capability, and uncertainty.