Logic-Aware Noise Injection (LNI)

• Logic-Aware Noise Injection (LNI) is a method that injects controlled noise constrained by domain-specific logic to maintain semantic integrity.
• It employs graph-level perturbations and tuned stochastic inputs in analog systems to enhance contrastive learning and Boolean reliability.
• The approach leverages logic-preservation constraints to improve adversarial robustness and optimize performance in complex network settings.

Searching arXiv for the cited papers to ground the article in current records. Logic-Aware Noise Injection (LNI) denotes a class of noise or perturbation mechanisms constrained by domain logic rather than applied arbitrarily. In the literature represented here, the term appears in two technically distinct settings. In provenance-based intrusion detection, LNI is a family of graph-level perturbation operators that generate semantically valid views of provenance graphs for contrastive learning in MirGuard, a robust anomaly detection framework against graph manipulation attacks (Sang et al., 14 Aug 2025). In an electronic analog of synthetic genetic networks, LNI functions as a design paradigm for injecting controllable noise into a bistable nonlinear system so that weak logic inputs are reliably converted into Boolean outputs within an optimal noise window, a phenomenon identified as Logical Stochastic Resonance (Hellen et al., 2012). Across both settings, the defining feature is that injected variability is constrained by the logic of the underlying system rather than treated as unrestricted corruption.

1. Terminological scope and core idea

In MirGuard, Logic-Aware Noise Injection is explicitly defined as a family of graph-level perturbation operators—edge augmentation (EA), node augmentation (NA), and feature augmentation (FA)—designed to inject noise into a provenance graph while preserving the graph’s “basic causal and type-level semantics” and yielding a plausible, logic-preserving view of the original graph (Sang et al., 14 Aug 2025). This distinguishes LNI from arbitrary structural perturbation. The perturbed graph is not intended to be a random corruption; it is intended to remain semantically valid under provenance-domain constraints.

In the earlier noise-aided logic setting, LNI is presented as a design paradigm extracted from an experimental study of an electronic analog of synthetic genetic networks (Hellen et al., 2012). There, the central idea is likewise non-arbitrary variability: noise is introduced into a nonlinear bistable system with weak logic inputs, adjustable biases, and thresholded outputs so that logic operations become reliable only within an intermediate noise regime. The paper characterizes this as noise-enhanced logic behavior and identifies the underlying mechanism as Logical Stochastic Resonance.

A common misconception is to equate LNI with indiscriminate noise addition. Both usages reject that interpretation. In MirGuard, the perturbation must satisfy an explicit logic predicate $L(u,v)$. In the analog logic system, the noise amplitude must be tuned to an optimal window; too little or too much noise degrades logical fidelity [(Sang et al., 14 Aug 2025); (Hellen et al., 2012)].

2. LNI in provenance graphs: formal model and operator family

MirGuard begins from an attributed, directed provenance graph

$G = (V, E, X, \tau_V, \tau_E),$

where $V = \{v_1,\dots,v_n\}$ is the set of nodes, $E \subseteq V \times V$ is the set of labeled directed edges, $\tau_V: V \to \mathcal{T}_V$ and $\tau_E: E \to \mathcal{T}_E$ are type-label functions for nodes and edges, and $X \in \mathbb{R}^{|V|\times d}$ is the matrix of initial one-hot-encoded node features (Sang et al., 14 Aug 2025). Nodes correspond to system entities such as processes, files, and networks; edges correspond to interactions such as read, write, connect, exec, and fork.

LNI is a randomized mapping

$$\mathcal{L}_\gamma: G \mapsto \tilde{G} = (\tilde{V}, \tilde{E}, \tilde{X}),$$

parameterized by augmentation ratios

$\gamma = (\gamma_E,\gamma_V,\gamma_X),$

which control the proportion of edges, nodes, or features that are to be added, removed, or swapped (Sang et al., 14 Aug 2025). The three operators are defined as follows.

Operator	Mechanism	Constraint
EA	Add and remove edges	Candidate edges must satisfy $L(u,v)=\top$
NA	Add and remove nodes, with incident-edge updates	New incident edges must satisfy the same logic check
FA	Replace a node feature with one from a same-type node	Require $G = (V, E, X, \tau_V, \tau_E),$0

For edge augmentation, two disjoint subsets are selected: $G = (V, E, X, \tau_V, \tau_E),$1 and $G = (V, E, X, \tau_V, \tau_E),$2, each of size $G = (V, E, X, \tau_V, \tau_E),$3, yielding

$G = (V, E, X, \tau_V, \tau_E),$4

Every candidate edge in $G = (V, E, X, \tau_V, \tau_E),$5 must satisfy the domain logic constraint $G = (V, E, X, \tau_V, \tau_E),$6 (Sang et al., 14 Aug 2025).

For node augmentation, a set $G = (V, E, X, \tau_V, \tau_E),$7 of new nodes of size $G = (V, E, X, \tau_V, \tau_E),$8 is created, each assigned a type $G = (V, E, X, \tau_V, \tau_E),$9, while a set $V = \{v_1,\dots,v_n\}$0 of size $V = \{v_1,\dots,v_n\}$1 is removed. All incident edges to removed nodes are deleted. Newly added nodes are connected to existing nodes via a small set of edges $V = \{v_1,\dots,v_n\}$2, each satisfying the same logic checks. The resulting graph satisfies

$V = \{v_1,\dots,v_n\}$3

with $V = \{v_1,\dots,v_n\}$4 formed from surviving original edges plus $V = \{v_1,\dots,v_n\}$5 (Sang et al., 14 Aug 2025).

For feature augmentation, for a fraction $V = \{v_1,\dots,v_n\}$6 of the nodes, the feature vector of node $V = \{v_1,\dots,v_n\}$7 is replaced by that of another node $V = \{v_1,\dots,v_n\}$8 of the same type:

$V = \{v_1,\dots,v_n\}$9

Equivalently, if $E \subseteq V \times V$0 of size $E \subseteq V \times V$1, then

$E \subseteq V \times V$2

while all other node features remain unchanged (Sang et al., 14 Aug 2025).

3. Semantic validity and logic-constraint enforcement

The defining technical property of MirGuard’s LNI is logic preservation. The predicate $E \subseteq V \times V$3 encodes allowed provenance interactions, and the paper specifies the admissible type-level combinations as Process$E \subseteq V \times V$4File (read/write), Process$E \subseteq V \times V$5Network (connect/send/recv), File$E \subseteq V \times V$6Process (exec/load), and Process$E \subseteq V \times V$7Process (fork/clone); all other combinations are logically forbidden, that is, $E \subseteq V \times V$8 (Sang et al., 14 Aug 2025).

During edge insertion in EA and during connection formation in NA, every candidate edge must pass $E \subseteq V \times V$9; otherwise it is rejected and a new candidate is drawn. This logic check is the mechanism that separates LNI from unconstrained graph perturbation. The paper’s illustrative mini-graph makes the point concretely: starting from $\tau_V: V \to \mathcal{T}_V$0 and $\tau_V: V \to \mathcal{T}_V$1, edge augmentation with $\tau_V: V \to \mathcal{T}_V$2 may add $\tau_V: V \to \mathcal{T}_V$3, which is allowed as exec, but would reject $\tau_V: V \to \mathcal{T}_V$4 (Sang et al., 14 Aug 2025).

The same example shows how node and feature augmentation remain type-consistent. With $\tau_V: V \to \mathcal{T}_V$5, the graph may remove $\tau_V: V \to \mathcal{T}_V$6, add a new node $\tau_V: V \to \mathcal{T}_V$7 of type process, and connect $\tau_V: V \to \mathcal{T}_V$8. With $\tau_V: V \to \mathcal{T}_V$9, the feature of $\tau_E: E \to \mathcal{T}_E$0 may be swapped with that of another process node. The paper states that each of these steps respects $\tau_E: E \to \mathcal{T}_E$1, $\tau_E: E \to \mathcal{T}_E$2, and the causal-semantic predicate $\tau_E: E \to \mathcal{T}_E$3 (Sang et al., 14 Aug 2025).

This logic-preserving requirement is central to the method’s interpretation. The perturbation is meant to simulate benign structural variance while excluding forbidden causal patterns. A plausible implication is that the induced augmentation distribution is narrower than that of generic graph corruption schemes, because candidate modifications are filtered through explicit provenance semantics.

4. Integration into logic-preserving contrastive learning

MirGuard uses LNI to form the positive pairs in a contrastive representation-learning pipeline (Sang et al., 14 Aug 2025). Two augmented graph views are constructed,

$\tau_E: E \to \mathcal{T}_E$4

possibly with different random seeds but the same logic rules. A GNN encoder $\tau_E: E \to \mathcal{T}_E$5 maps these views to graph-level embeddings $\tau_E: E \to \mathcal{T}_E$6 and $\tau_E: E \to \mathcal{T}_E$7, after which a two-layer projector $\tau_E: E \to \mathcal{T}_E$8 yields $\tau_E: E \to \mathcal{T}_E$9 and $X \in \mathbb{R}^{|V|\times d}$0.

The contrastive objective is

$X \in \mathbb{R}^{|V|\times d}$1

Because the positive pair $X \in \mathbb{R}^{|V|\times d}$2 is generated from logic-aware views, the model is encouraged to learn representations that are invariant to benign, logic-preserving transformations but not invariant to adversarial inconsistencies (Sang et al., 14 Aug 2025). The paper explicitly argues that adversarial manipulations that break causality, such as forbidden edges, will not appear as positives, thereby sharpening sensitivity to such inconsistencies at inference time.

The empirical argument for this design has two parts. First, the paper reports that MirGuard’s F1 and AUC under five classes of graph-manipulation attacks degrade far less than baselines, and attributes this to the model being trained to ignore logic-preserving perturbations while detecting logic-violating ones (Sang et al., 14 Aug 2025). Second, the ablation results state that removing EA, NA, or FA reduces robustness, and that replacing the contrastive backbone with DGI or GraphSAGE also drops performance. The paper interprets this as demonstrating that the combination of logic-aware augmentation plus contrastive loss is necessary and sufficient for the observed adversarial resilience (Sang et al., 14 Aug 2025). It also characterizes MirGuard as the first targeted study to enhance provenance-based intrusion detection systems against graph manipulation attacks.

5. Hyperparameters, operating regime, and implementation logic

MirGuard specifies three principal augmentation hyperparameters. The edge-level perturbation ratio $X \in \mathbb{R}^{|V|\times d}$3 controls the intensity of edge addition and removal. According to the paper, values that are too small, written as $X \in \mathbb{R}^{|V|\times d}$4, fail to expose the model to meaningful structural variance, whereas values that are too large, written as $X \in \mathbb{R}^{|V|\times d}$5, may distort global semantics. The empirical optimum is reported as approximately $X \in \mathbb{R}^{|V|\times d}$6 (Sang et al., 14 Aug 2025).

The node-level augmentation ratio $X \in \mathbb{R}^{|V|\times d}$7 governs insertion and deletion of nodes, with typical values listed as $X \in \mathbb{R}^{|V|\times d}$8 or $X \in \mathbb{R}^{|V|\times d}$9. The feature-level ratio $$\mathcal{L}_\gamma: G \mapsto \tilde{G} = (\tilde{V}, \tilde{E}, \tilde{X}),$$0 determines the fraction of feature swaps, and the paper states that FGPA robustness peaks at moderate $$\mathcal{L}_\gamma: G \mapsto \tilde{G} = (\tilde{V}, \tilde{E}, \tilde{X}),$$1. The choice of the logic predicate $$\mathcal{L}_\gamma: G \mapsto \tilde{G} = (\tilde{V}, \tilde{E}, \tilde{X}),$$2 is fixed by domain semantics (Sang et al., 14 Aug 2025).

The pseudocode given for LNI makes the implementation discipline explicit. For EA, the method samples edge candidates uniformly from $$\mathcal{L}_\gamma: G \mapsto \tilde{G} = (\tilde{V}, \tilde{E}, \tilde{X}),$$3 until the target number of logic-valid additions is reached, then samples removals uniformly from existing edges. For NA, it samples a removal set, creates new typed nodes, and connects each added node by at least one logic-valid edge. For FA, it samples nodes uniformly and replaces their features by those of same-type nodes. The final perturbed graph is then assembled from the surviving and added components (Sang et al., 14 Aug 2025).

A recurring point in the paper is that robustness is not obtained by perturbation magnitude alone. The perturbation must remain semantically plausible. This directly counters the intuition that adversarial robustness in graph learning can be achieved by stronger random augmentation irrespective of domain constraints.

6. Noise-aided logic in synthetic genetic network analogs

In the electronic analog of synthetic genetic networks, the system evolves in dimensionless time $$\mathcal{L}_\gamma: G \mapsto \tilde{G} = (\tilde{V}, \tilde{E}, \tilde{X}),$$4 with dimensionless concentrations $$\mathcal{L}_\gamma: G \mapsto \tilde{G} = (\tilde{V}, \tilde{E}, \tilde{X}),$$5 and $$\mathcal{L}_\gamma: G \mapsto \tilde{G} = (\tilde{V}, \tilde{E}, \tilde{X}),$$6, and the dynamics include Hill-type inhibition, additive biases, external inputs, and additive noise (Hellen et al., 2012). The parameters $$\mathcal{L}_\gamma: G \mapsto \tilde{G} = (\tilde{V}, \tilde{E}, \tilde{X}),$$7 are maximal expression rates, $$\mathcal{L}_\gamma: G \mapsto \tilde{G} = (\tilde{V}, \tilde{E}, \tilde{X}),$$8 are effective decay rates, $$\mathcal{L}_\gamma: G \mapsto \tilde{G} = (\tilde{V}, \tilde{E}, \tilde{X}),$$9 is the Hill coefficient, $\gamma = (\gamma_E,\gamma_V,\gamma_X),$0 are constant bias currents, $\gamma = (\gamma_E,\gamma_V,\gamma_X),$1 are low-amplitude logic inputs, and $\gamma = (\gamma_E,\gamma_V,\gamma_X),$2 is the noise strength. The noise terms $\gamma = (\gamma_E,\gamma_V,\gamma_X),$3 are independent unit-variance, zero-mean white Gaussian processes.

The logic mechanism is Logical Stochastic Resonance. At zero or very low noise, the system tends to remain in whichever well it started in, so weak logic inputs cannot reliably flip the state. At very large noise, spontaneous barrier crossings become frequent and random, and output loses correlation with inputs. At an intermediate sweet spot $\gamma = (\gamma_E,\gamma_V,\gamma_X),$4, noise assists the barrier crossings triggered by the input pulses, and thresholding $\gamma = (\gamma_E,\gamma_V,\gamma_X),$5 or $\gamma = (\gamma_E,\gamma_V,\gamma_X),$6 produces Boolean outputs that reproduce AND and the complementary NAND with near-unity fidelity (Hellen et al., 2012).

The paper defines an accuracy metric

$\gamma = (\gamma_E,\gamma_V,\gamma_X),$7

where $\gamma = (\gamma_E,\gamma_V,\gamma_X),$8 is the fraction of correct outputs for the low-output cases of the AND truth table and $\gamma = (\gamma_E,\gamma_V,\gamma_X),$9 is the fraction of correct outputs for the $L(u,v)=\top$0 case. Both numerical simulation and oscilloscope-based measurement yield a bell-shaped $L(u,v)=\top$1 curve peaking near $L(u,v)=\top$2, which the paper identifies as the hallmark of Logical Stochastic Resonance (Hellen et al., 2012).

The circuit implementation uses two identical gene-modules, each with a pnp transistor for protein production, a resistor-capacitor integrator, and cross-coupling so that each gene’s output represses the other. Inputs $L(u,v)=\top$3 and $L(u,v)=\top$4 are implemented by additional pnp transistors, and noise is injected into each gene’s input node through a resistor connected to a two-stage noise amplifier based on a reverse-biased B–E junction of a 2N3904 and cascaded OPA2228 op-amps. The resulting noise is approximately white over $L(u,v)=\top$5 MHz. Outputs are read as dimensionless voltages and thresholded at $L(u,v)=\top$6 for Boolean readout (Hellen et al., 2012).

The same hardware can be morphed between logic families. With $L(u,v)=\top$7 and small inputs $L(u,v)=\top$8, the circuit realizes AND on $L(u,v)=\top$9 and complementary NAND on $G = (V, E, X, \tau_V, \tau_E),$00. Introducing a small constant bias $G = (V, E, X, \tau_V, \tau_E),$01 tilts the potential so that a single active input suffices to trigger switching, yielding OR on $G = (V, E, X, \tau_V, \tau_E),$02 and NOR on $G = (V, E, X, \tau_V, \tau_E),$03 (Hellen et al., 2012).

7. Comparative interpretation, misconceptions, and limits

The two uses of LNI share a common structural idea: variability is introduced in a way that respects the logic of the target system. In MirGuard, the relevant logic is causal and type-level validity in provenance graphs. In the synthetic genetic network analog, the relevant logic is the mapping from weak input pulses, bistable dynamics, and thresholded outputs to Boolean truth tables [(Sang et al., 14 Aug 2025); (Hellen et al., 2012)].

They differ, however, in what “noise” means operationally. In MirGuard, noise is a family of constrained graph transformations used to generate positive pairs for contrastive learning. In the analog system, noise is an additive stochastic drive whose amplitude must be tuned to an intermediate regime. This suggests that “Logic-Aware Noise Injection” functions less as a single standardized algorithm than as a general principle: perturbation or stochasticity should be admissible under the semantics of the system being modeled.

Several misconceptions are directly addressed by the source material. First, LNI is not arbitrary augmentation or arbitrary hardware noise. MirGuard rejects edge and node changes that violate $G = (V, E, X, \tau_V, \tau_E),$04, and the analog logic study shows that logic fidelity fails both when noise is too weak and when it is too strong [(Sang et al., 14 Aug 2025); (Hellen et al., 2012)]. Second, robustness claims in MirGuard are not presented as deriving from a long formal proof; the paper explicitly states that the justification is empirical and ablation-based. Third, the analog work does not claim monotonic benefit from increased noise; it reports an optimal window and a bell-shaped accuracy profile.

Taken together, these two strands define LNI as a logic-constrained approach to stochasticity or perturbation. In one case it is a provenance-graph augmentation mechanism for adversarially robust representation learning; in the other it is a blueprint for reliable, reconfigurable Boolean computation in analog nonlinear hardware.