Variance-Driven Mirage: Impact on Systems & ML

• Variance-Driven Mirage is a phenomenon where neglecting true randomness leads to misinterpreted correlations in security, machine learning, and scheduling systems.
• In randomized caching, fixed RNG seeds artificially boost side-channel signals, while proper per-run randomness destroys these misleading patterns.
• Empirical studies in reinforcement learning and cluster scheduling underscore that variance-aware modeling is essential for avoiding flawed performance gains and design errors.

Variance-Driven Mirage refers to the phenomenon in security, systems, and machine learning domains where the introduction or accurate modeling of variance fundamentally alters the interpretation of signal, causality, or attack feasibility. Across contexts, variance-driven mirages manifest when empirical or simulated results suggest reliable correlations or performance, but those effects vanish once sources of randomness or uncertainty are properly incorporated. Thus, what appears as actionable signal is revealed to be either a modeling artifact or an implementation bias, not a robust property of the underlying system.

1. MIRAGE in Randomized Caches: The Eviction Variance Mechanism

In the domain of hardware security, Mirage is a fully-associative randomized caching scheme that aims to disrupt side-channel leakage by randomizing cache line evictions each time an encryption operation is performed. The system models the cache as having capacity $K$ lines. At each encryption, a sequence of victim table-lookups $V_t$ is made, and every cache miss or installation triggers a random eviction: one of the $K$ resident lines is evicted with uniform probability, dictated by a random number generator (RNG) seeded per encryption run.

Key phenomena arise when analyzing attacker models:

• Constant-seed simulation: If one erroneously fixes the RNG seed ($seed_t = S_0$ for all $t$), all eviction sequences $E_t$ are identical across runs. This creates deterministic timing profiles, highly correlated with victim table accesses. Occupancy-based attacks then appear effective.
• Random-seed simulation: In the realistic scenario ($seed_t$ i.i.d. uniform), eviction sequences $E_t$ vary independently per run, sampling uniformly from all possible eviction permutations.

When attackers perform correlation analysis—e.g., computing Pearson $\rho_n(k)$ between key guesses and observed timing—the constant-seed case exhibits artificially high correlation (signal-to-noise ratio $S/N$ is large). However, with randomized evictions, the non-repeatable, stochastic eviction pattern introduces $V_t$0 additional variance (for $V_t$1 attacker probes, cache capacity $V_t$2), causing $V_t$3 to collapse toward zero ($V_t$4, typically $V_t$5 for practical parameters), indistinguishable from background noise (Cao et al., 14 Aug 2025).

2. Correlation, Entropy, and Empirical Disappearance of Signal

Empirical validation underscores this phenomenon. Heatmaps of attacker access time vs. T-table index in constant-seed scenarios show clear, information-theoretic structures (low guessing entropy, high classification accuracy of AES keys). Once per-run seed variance is introduced (e.g., calling srand(time()) before each run), this structure vanishes: the heatmaps display no key-dependent correlation, and guessing entropy remains at or above 90% even for thousands of runs. Raw histograms of probe timings collapse into broad, overlapping distributions without discriminatory power.

The implication is that any attack claiming to exploit occupancy-based signal in MIRAGE is merely observing the deterministic outcome of fixed-seed simulation. True hardware operation, incorporating fresh eviction variance per invocation, destroys the side-channel signal. The apparent vulnerability is thus a variance-driven mirage—a modeling illusion, not a genuine leakage (Cao et al., 14 Aug 2025).

3. Reinforcement Learning: Mirage of Variance Reduction

Variance-driven mirages also arise in reinforcement learning, particularly within policy-gradient methods employing state-action-dependent baselines. Certain works posited that these baselines could substantially lower estimator variance without introducing bias, with resultant gains in sample efficiency. However, a decomposition of total estimator variance reveals the following:

$V_t$6

Here, only the second term ($V_t$7) is affected by the baseline. While the optimal control-variate baseline can eliminate $V_t$8, empirical studies show that, once a state-dependent baseline is already in use, $V_t$9 is negligible compared to trajectory-induced variance $K$0. Learning additional state-action dependence yields no pragmatic variance reduction; observed empirical gains were traced to subtle implementation biases, such as asymmetric normalization or overfitting the baseline on reused data, not true algorithmic improvement (Tucker et al., 2018).

4. Variance Sensitivity in Batch Scheduling Systems

In cluster resource provisioner settings, as exemplified by the Mirage proactive provisioner for Slurm-compatible batch GPU clusters, performance is sensitive to variance in system workload and trace noise. The RL-driven agent’s ability to reduce interruptions (and overlap) relies on accurately identifying and reacting to stochastic job arrival and queue dynamics. Transformer-based PG and DQN models embedded in Mirage demonstrate varying robustness to arrival variance:

• Transformer+PG: Yields lowest interruptions on heavy workloads but is sensitive to noisy traces, sometimes introducing excessive overlap in light workloads.
• MoE+DQN: More robust to stochastic, noisy job arrivals, preferred for balanced operational trade-offs.

A plausible implication is that design of resource-scheduling policies must actively acknowledge and optimize variance both in reward and system state—variance-unaware methods may exhibit apparent performance only under specific, less stochastic regimes, while performance collapses in truly variable operational contexts (Ding et al., 2023).

5. Common Mechanisms and Theoretical Underpinnings

Variance-driven mirage conditions share several mechanistic features:

Scenario	Source of Mirage	Effect of Accurate Variance Modeling
MIRAGE Side-Channel Attack	Fixed RNG seed per run	Signal destroyed, attack fails
RL Policy Gradient Baseline	Unaccounted bias/overfit	Empirical variance reduction vanishes
Batch Scheduling (Mirage)	Trace/arrival variance	RL model performance variable

In all cases, the mirage results from deterministic structure unrepresentative of the real system, erased once stochasticity is properly modeled.

6. Implications, Limitations, and Future Directions

Variance-driven mirages illustrate the necessity of faithful system modeling, particularly regarding the initialization and propagation of randomness. Their identification is critical for validating attack models, estimator improvements, and resource allocation policies. In caching, faithful per-run seed variance is essential; in RL, decomposition of estimator variance and vocabulary of unbiased baselines must be rigorously adhered to; in scheduling, operational robustness hinges on variance-aware design.

Potential future extensions include explicit variance-aware objectives for resource provisioners, hybrid reward structures balancing mean and variance penalties, and system design embracing transfer learning to adaptively mitigate impact of stochasticity across operational transitions (Ding et al., 2023). Recognition and correction of variance-driven mirages remain central to advancing security, control, and systems research.