Infector-Omission (IO): Mechanisms and Impact

• Infector-Omission (IO) is the phenomenon where pivotal nodes or elements are omitted, disrupting propagation, observability, and system robustness across various domains.
• In software testing, ML, control theory, and epidemiology, IO disrupts detection and propagation by omitting influential components, leading to undetected mutations, misclassifications, filtering disturbances, and tracing gaps.
• Mitigation strategies include enhancing test coverage, adopting robust ML training, utilizing advanced filtering techniques, and ensuring comprehensive tracing to counteract IO’s detrimental effects.

Infector-Omission (IO) denotes scenarios in which particular nodes, elements, or steps in a system are systematically omitted, resulting in distinct propagation, observability, or robustness phenomena. The IO concept is found in diverse domains, including software testing (mutation analysis), adversarial data attacks in machine learning, robust filtering in control theory, and epidemiological modeling. Though domain-specific, IO consistently refers to the omission of a causative element (e.g., an infector, influential data point, or state-update), thereby disrupting propagation or detection mechanisms.

1. Formal Definitions and Contexts

Software Testing: No-Infection Symptom

In mutation analysis, IO corresponds to the "no-infection" condition during extreme transformation testing. Consider a method $m$ and its variant $m'$ (in which $m$'s body is emptied). Let $T$ be the set of test cases and $T_m \subseteq T$ those that cover $m$. Define the observed immediate program state as $MS_{m,t}$ and $MS_{m',t}$ for $m$ and $m'$ respectively. IO holds if

$m'$0

or equivalently, using a difference function

$m'$1

No test in $m'$2 causes any observable change at the mutation site; the transformation remains undetectable (Vera-Pérez et al., 2019).

Machine Learning: Targeted Data-Sample Omission Attacks

IO-style attacks are formalized as the targeted omission of training data to influence model behavior. Given a training set $m'$3 and target example $m'$4, the attacker selects an omission set $m'$5, $m'$6, such that training on $m'$7 leads to misclassification of $m'$8. The optimization can be expressed as

$m'$9

where $m$0 is the estimated model and $m$1 the loss (Barash et al., 2021).

Robust Filtering: Innovation Outliers

In state-space models,

$m$2

IO signifies "innovation outliers"—large, propagating disturbances in the state equation. The contaminated increment is

$m$3

with $m$4 governing outlier occurrence. Robust filters are then designed to minimize worst-case mean squared error under IO contamination (Ruckdeschel, 2010).

Epidemiology: Infector-Omission in Contact Tracing

In stochastic agent-based epidemic modeling, IO is the fraction $m$5 of infectors not traced by contact tracing (CT). Let $m$6 indicate tracing of infector $m$7: $m$8 All edges (contacts) from an omitted infector (where $m$9) are invisible to CT. The effective next-generation matrix becomes $T$0, with corresponding reproduction number $T$1, where $T$2 is the fraction of transmissions prevented by perfect tracing (Chae et al., 21 Jan 2026).

2. Algorithmic and Analytical Approaches

Detection and Diagnosis (Software Testing)

Reneri implements IO identification via a two-pass dynamic analysis:

• Infection-Detection Pass: Instrument both $T$3 and $T$4, observe the local state $T$5 and $T$6 across $T$7, compute the state difference, and label transformations as IO if no difference is recorded.
• Classification Algorithm: Partition undetected transformations into no-infection (IO), no-propagation, and weak-oracle symptoms using state-diff and test-case-level state propagation (Vera-Pérez et al., 2019).

Omission Set Construction (ML Attacks)

Greedy and Genetic heuristic algorithms are deployed:

• Greedy: Iteratively removes training points that most increase loss on the target $T$8.
• Genetic: Evolves populations of omission sets to maximize misclassification probability, using cross-over and mutation within a fixed budget $T$9 (Barash et al., 2021).

Robust Kalman Filtering

For IO robustness, a minimax filter modifies the standard Kalman correction:

1. Predict new state estimate.
2. Compute innovation and Kalman gain.
3. Apply a Huber-type clipping to innovation residuals to attenuate outlier impact.
4. Update state and covariance recursively (Ruckdeschel, 2010).

Epidemic Modeling under IO

Simulations implement $T_m \subseteq T$0-based removal of infector trajectories per $T_m \subseteq T$1 at each confirmation event, modifying the traced adjacency matrix for downstream quarantine and transmission analysis (Chae et al., 21 Jan 2026).

3. Empirical Evidence and Phenomenological Patterns

Software Testing Prevalence and Patterns

Out of 312 undetected extreme transformations across 15 Java projects:

• 114 (37%) showed IO.
• For boolean-returning methods, 65% of such transformations exhibited IO.
• Precondition checks frequently generated IO due to lack of comprehensive input coverage.
• IO is not rare; it indicates that the input space never causes any observable infection at the targeted location (Vera-Pérez et al., 2019).

Machine Learning Attack Success

Observed IO attack (targeted data-sample omission) success rates:

• White-box: up to 100% for some learners/datasets (MNIST, ANN, Genetic algorithm).
• Black-box: success rates up to 85% (MNIST, $T_m \subseteq T$2-NN).
• Model accuracy typically degrades by only 1–3% with omission budgets $T_m \subseteq T$3 of $T_m \subseteq T$4 (Barash et al., 2021).

Kalman Filtering Performance

In simulation:

• Under pure IO contamination (innovation outliers), the rLS.IO filter achieves lowest mean squared error (MSE ≈ 1.05), outperforming classical Kalman (MSE = 3.12) and AO-robust filters (MSE = 48.3).
• With mixed IO+AO outliers, hybrid or nonparametric filters offer better robustness (Ruckdeschel, 2010).

Epidemiological Impact

In simulated epidemics (Seoul, Busan):

• For Seoul, $T_m \subseteq T$5 triggers an abrupt increase in cumulative infections from <1% to ≈15–20% of the population.
• Mean network diameter grows dramatically (from ~10 to ~50) as $T_m \subseteq T$6 increases, reflecting deeper and longer transmission chains.
• Busan exhibits a higher tolerance with threshold at $T_m \subseteq T$7, attributed to lower density and differing contact structure (Chae et al., 21 Jan 2026).

Domain	IO Manifestation	Empirical Impact/Rate
Software Test	No-infection mutants	37% undetected mutations
ML Attacks	Targeted omission	80–100% success (white-box)
Control	Innovation outlier	MSE↓ (>3× vs. classical KF)
Epidemiology	Untagged infectors	Epidemic threshold at ≈4%

4. Practical Implications and Remediation

Test Suite Improvements

Reneri issues actionable suggestions for IO: augment the test input space so that mutated method invocations can infect some observable state component. This involves generating new test cases or modifying arguments so that output or side effects differ when compared to the emptied variant (Vera-Pérez et al., 2019).

Adversarial Defense in Learning

Potential defenses against IO attacks:

• Robust/safe training incorporating regularization (e.g., influence-function penalties).
• Auditing and provenance requirements for data deletion.
• Model ensembling for cross-class invariance (Barash et al., 2021).

Outlier-Robust Signal Processing

Robust IO filters implement innovation clipping in recursive state estimation, trading minor ideal-case efficiency for significant resistance to propagating outliers. Heuristic switching or hybridization with AO-robust designs extends coverage to mixed fault scenarios (Ruckdeschel, 2010).

Epidemic Policy

Given the abrupt infection threshold under IO, maximizing successful infector tracing (minimizing $T_m \subseteq T$8) is critical for containment. Small increases in IO yield drastic reductions in CT efficacy; priority should be on tracing completeness over contact granularity (Chae et al., 21 Jan 2026).

5. Comparative Perspective: IO Versus Related Phenomena

Infector-Omission consistently denotes omission of causative elements, in contrast to “contact-omission” (epidemiology: missed edges), “additive outlier” (AO, filtering: observation noise), or “propagation failure” (software: infection not observed by oracles). Unlike non-propagating omissions, IO is characterized by its impact on upstream propagation and subsequent system observability or controllability.

In software, IO is the root cause of certain undetectable mutants; in ML, omission is exploited adversarially; in control, IO corresponds to propagating faults; and in epidemics, IO dramatically undermines tracing effectiveness.

6. Limitations, Open Questions, and Outlook

IO phenomena expose fundamental structural vulnerabilities across domains:

• In software, IO typically signals inadequate input domain coverage rather than oracle weakness.
• In machine learning, the extent of influence of a data point—and the model class’s capacity—limits the attack’s power.
• In control, ideal IO-robust filters are less performant under AO or mixed outliers, inviting further hybrid methodological development.
• In epidemiology, the abruptness of the IO threshold suggests unknowns regarding subnetwork resilience, heterogeneity in infector connectivity, and the value of partial tracing.

A plausible implication is that IO analysis provides both a diagnostic tool for process deficiencies (software, CT) and a design challenge for robustness in adversarial and noisy scenarios (ML, signal processing). Cross-domain synthesis—translating insights and mitigation techniques—remains an open research avenue.