Conditional Process Models of Side-Channel Leakage: Locating Where Defenses Actually Intervene

When software runs on someone else's hardware, encryption does not by itself hide what the software is doing. The sequence of memory addresses a program touches — which cache lines it evicts, which pages it faults on, whether each operation is a read or a write — reveals a great deal about the data being processed. Attacks exploit this: adversaries have reconstructed human faces from a stream of 1-bit read/write flags Private Image Reconstruction from System Side Channels Using Generative Models, identified websites from cache occupancy on Apple Silicon EXAM: Exploiting Exclusive System-Level Cache in Apple M-Series SoCs for Enhanced Cache Occupancy Attacks, and broken address-space randomization on AMD confidential VMs by exploiting data sparsity Relocate-Vote: Using Sparsity Information to Exploit Ciphertext Side-Channels. The standard defensive answer — making every memory access data-independent ("obliviousness") — is provably safe but costs 16.7x to 178.2x overhead in realistic enclave workloads Secure Collaborative Training and Inference for XGBoost, so almost nobody deploys it. What practitioners actually run is partial, messy protection, and the field has no language for reasoning about it beyond "not formally secure."

This proposal imports a statistical tool from the social sciences — conditional process analysis, also called moderated mediation — and argues it is exactly the missing language. A moderated mediation model takes a causal chain $X \to M \to Y$ (a driver, a mediator, an outcome) and asks not just how much effect flows through the chain, but how that flow changes with a boundary condition $W$ College students' cyberloafing and the sense of meaning of life: The mediating role of state anxiety and the moderating role of psychological flexibility. Applied here, $X$ is a sensitive property of the data or program, $M$ is the microarchitectural observable the host can see (cache collisions, fault counts, trace statistics), $Y$ is what the adversary actually extracts, and $W$ is a system condition or defense parameter. The payoff is a way to say, quantitatively, whether a defense stops the leak from being generated in the first place or merely degrades the adversary's decoding of it, and whether a defense that suppresses one channel silently amplifies another.

Core Claim

Side-channel leakage is a conditional process, and modeling it as one changes what defenses we can justify. Concretely, I claim three things. First, the transmission chain from secret ($X$) through observable trace ($M$) to adversary success ($Y$) can be decomposed into path coefficients $a$ (secret to trace) and $b$ (trace to success), with a conditional indirect effect $\omega(W) = (a_1 + a_2 W)\,b$ when a system condition $W$ moderates the first stage — the standard machinery of conditional process models The relationship between family functioning and social media addiction among university students: a moderated mediation model of depressive symptoms and peer support. Second, defenses can be classified by which link they moderate: first-stage moderators (e.g., oblivious sorting that suppresses signal generation) versus second-stage moderators (e.g., adversary-facing noise like Apple's cache-masking stride, which costs under 5% overhead for single-core tasks EXAM). Third — and this is where the multi-path machinery earns its keep — a single defense can be a "double-edged sword," exactly as peer support proved to be in Qi et al.: partial obliviousness buffers the access-pattern channel while its added traffic and runtime may amplify write-volume and timing channels Connecting the Dots: Privacy Leakage via Write-Access Patterns to the Main Memory, High Recovery with Fewer Injections: Practical Binary Volumetric Injection Attacks against Dynamic Searchable Encryption. No existing work measures this trade-off as a joint conditional process; the closest papers observe its ingredients separately. Relocate-Vote shows sparsity modulating ciphertext leakage but treats it as a static domain property with no path decomposition; EXAM shows GPU compression and cache inclusivity acting as physical moderators but reports ablation tables, not conditional indirect effects; the oblivious-differential-privacy line An Algorithmic Framework For Differentially Private Data Analysis on Trusted Processors., Oblivious Sampling Algorithms for Private Data Analysis bounds worst-case leakage but offers no empirical, continuous model of how operational conditions attenuate or amplify it. A Lacuna search confirms no work applies conditional process analysis to hardware side channels; retrieved "conditional memory" papers concern LLM attention routing, not leakage.

A methodological point makes this cross-domain import worth doing rather than decorative: the biggest weakness of conditional process analysis in its home domain is causal ambiguity — cross-sectional self-report surveys cannot establish direction or rule out confounders, as Qi et al. themselves acknowledge. Hardware testbeds have ground truth and, crucially, experimental control: we can randomize inputs, defenses, and environmental conditions, so the estimated conditional indirect effects are experimentally identified, not merely correlational. The import thus fixes a limitation in both directions.

Research Program

The work has three deliverables. First, a measurement harness that instruments existing attack pipelines to record the mediator layer explicitly: for each experimental run we log the secret ($X$: image identity, page type, ASLR layout), a vectorized observable ($M$: address-histogram entropy, page-fault counts, trace length, read/write bit-sequence statistics, cache-set collision counts), and the adversary outcome ($Y$: matching accuracy, search-space reduction, reconstruction quality). Existing attacks already log $X$ and $Y$ but discard $M$ or bury it inside an end-to-end neural pipeline, which is precisely what prevents mechanism claims. Second, a fitting procedure: two-stage path regression with moderator interactions, bootstrapped confidence intervals on $\omega(W)$, following the established conditional process estimators used in Qing Li et al. (Model 7, single-stage moderation) and Qi et al. (Model 59, multi-path moderation), extended to two parallel mediators (access pattern and volume/timing) so we can express net conditional leakage as $\Lambda(W) = \omega_{\text{access}}(W) + \omega_{\text{vol}}(W)$. Third, a design tool: a "conditional leakage profile" that plots $\Lambda(W)$ against overhead $C(W)$ for each defense configuration, letting a deployer choose the knee of the curve instead of choosing between 5x and 178x.

Proposed Experiments

Experiment 1 (reconstruction testbed) reproduces the generative attack of Yuan et al. on CelebA traces (700K–1.3M records per trace, down to the 1-bit read/write setting), sweeping moderators $W$: trace length, image pool size, library version (with and without chrominance-dependent accesses), and background cache contention. The decisive retrodiction test is their own observed color blindness: the framework should recover, from the coefficients, that chrominance transmission dies at the first stage (libraries lack color-dependent accesses ⇒ $a_2 \approx 0$ for that path) while luminance transmission survives — something their CNN–VAE–GAN pipeline observed but could not explain. Experiment 2 (defense trade-off testbed) re-instruments the enclave XGBoost setting of Andrew Law et al. on the Ant Financial dataset (100,000 samples, 126 features), varying the degree of obliviousness (fraction of dummy nodes, linear scans vs. bitonic sorting) and measuring both access-pattern fingerprinting accuracy and write-volume leakage of the kind demonstrated by Connecting the Dots. The target result is a configuration whose $\Lambda(W)$ is statistically indistinguishable from the full-obliviousness floor while its overhead sits below the 16.7x low end of the oblivious range. Experiment 3 (prediction test) evaluates whether fitted coefficients from a sparse grid of $W$ values predict attack success at held-out $W$ values, against the baselines below. As a secondary lens, EXAM's Alexa Top 100 website-identification setting EXAM provides a second domain for cross-testbed validation of the moderator taxonomy (Performance vs. Efficiency cores, background SLC saturation).

Baselines

The naive reference point is the do-nothing measurement: run the attack end-to-end and report accuracy, the way EXAM reports 90.5% website identification on the Alexa Top 100 and Relocate-Vote reports collapsing the glibc ASLR search space from over 8 million candidates to 35–104 with 9–10 successes in 10 runs. This baseline answers "is the system vulnerable here" but cannot say where an intervention acts, which is the question this proposal targets. The strongest existing competitor is the end-to-end deep generative pipeline of Yuan et al.: CNN trace featurization, a VAE with learned prior, and a GAN refiner, which reports 90.7% top-20 matching accuracy on CPU cache traces from a 100-image CelebA pool. We will run it on identical traces, compute, and tuning budget, and beat it in two specific senses: predicting its success at held-out moderator values with lower error than a per-condition refit (the process model has far fewer parameters, so it should extrapolate where the generator must retrain), and correctly diagnosing which stage fails when it fails. The third baseline is aggregate mutual information, $I(X;M)$ and $I(X;\text{trace})$ computed on the same runs — the standard information-theoretic summary. The fourth is the ablation that isolates the proposal's key ingredient: a plain mediation model with no moderation (fixed-$W$ path decomposition), and a moderation-only interaction regression with no mediator. If the full conditional-process structure adds nothing over these ablations, the central claim is wrong and I will say so.

Why not just…?

Why not just use mutual information or min-entropy? Aggregate information measures give channel capacity, not mechanism. The file's own evidence shows why that is insufficient: Yuan et al. proved a nominally negligible 1-bit read/write channel suffices to reconstruct recognizable faces once a generative prior mediates decoding — an aggregate mutual-information budget would dismiss exactly the channel that works. Mutual information also cannot express "this defense suppresses the $a$ path while this one suppresses the $b$ path," which is the actionable output. Experiment 3 settles this directly: if MI profiles predict held-out attack success as well as the conditional process model, the extra structure is not paying rent.

Why not just train an end-to-end deep model? Because the goal is defense design, not attack maximization. The Yuan et al. pipeline is the best extractor we know of, and we adopt it as the $M \to Y$ stage; but it is opaque about mechanisms — its authors could observe that colors were missing but not localize why. The conditional process model wraps such models rather than replacing them, and adds what they cannot give: stage attribution, moderation profiles, and extrapolation to unseen conditions without retraining. If it fails to out-predict the refit generator in Experiment 3, this objection wins.

Why not just run a regression with interaction terms, or an ANOVA? A single regression conflates direct and mediated transmission. The file's clearest illustration is Qi et al.: family functioning had no significant direct effect on addiction once depressive symptoms were included, with the entire effect flowing through an indirect path of $-0.22$ (95% CI $[-0.28, -0.17]$) — an interaction-only analysis would have wrongly concluded the driver was irrelevant. In our setting, an architectural feature may show no direct association with extracted secrets while strongly driving the observable trace; mediation decomposition is what separates these, and Experiment 2's defense trade-off (access leakage falling while volume leakage rises) is precisely the case a single-equation model averages away.

Why not just require full obliviousness or ODP guarantees? This is the strongest objection, and the honest answer is that formal guarantees remain the gold standard for high-stakes deployments — this proposal does not replace them. But the file documents that worst-case protection is frequently unaffordable (up to 178.2x overhead in Law et al.) and that theoretical guarantees can silently fail on real hardware: mathematically sound differential-privacy amplification degraded from $\epsilon = 0.82$ to $\epsilon = 9.39$ on DP-MNIST purely because the host observed access patterns during sampling Oblivious Sampling Algorithms for Private Data Analysis. For the deployments that will never pay 178x, the choice is not "oblivious vs. broken" but which partial protection to buy; the conditional leakage profile is the instrument for that decision, and it also empirically probes where worst-case assumptions break — a complement to, not a substitute for, the ODP program Allen et al..

Decisive Checks

Three results would make or break the proposal. First, the retrodiction: the fitted model must recover the color-blindness boundary in the CelebA testbed as a first-stage moderation ($a_2 \approx 0$ for chrominance), matching what Yuan et al. observed empirically. Second, the trade-off discovery: on the enclave XGBoost testbed, the framework must find at least one configuration whose net conditional leakage $\Lambda(W)$ is within the full-obliviousness floor's confidence interval while overhead stays under 16.7x over the encrypted baseline. Third, the prediction win: coefficients fitted on a sparse grid of moderator values must predict held-out attack success (CelebA matching accuracy, EXAM website identification) with meaningfully lower error than the mutual-information baseline and the per-condition-refit generator. If any of these fails, the fallback contribution is still a measured, multi-path account of the access-versus-volume defense trade-off, which the literature currently asserts but has never jointly quantified.