Granger Causality

The foundational principle of Granger causality: cause must precede effect. For collusion detection, this means Agent A's suspicious action at time t must consistently occur before Agent B's correlated action at time t+1. The test evaluates whether past values of Agent A's action sequence provide statistically significant information about Agent B's future actions, beyond what Agent B's own history already explains. This temporal ordering distinguishes genuine influence from mere correlation.

Granger causality is implemented through a Vector Autoregression (VAR) model that captures linear interdependencies among multiple time series. In a two-agent system, two equations are estimated:

• Agent B's actions regressed on its own lagged values and lagged values of Agent A
• Agent A's actions regressed on its own lagged values and lagged values of Agent B An F-test then determines if the lagged coefficients of the suspected causal agent are jointly statistically significant. A significant result implies Agent A 'Granger-causes' Agent B.

A critical precondition for valid Granger causality testing is stationarity—the statistical properties of the time series (mean, variance, autocorrelation) must remain constant over time. Non-stationary agent behavior data, such as trending transaction volumes or seasonal patterns, can produce spurious regression results that falsely indicate causality. Preprocessing steps like differencing or log transformation are applied to stabilize the series before testing. The Augmented Dickey-Fuller (ADF) test is commonly used to verify stationarity.

The choice of lag length—how many historical time steps to include—directly impacts test sensitivity. Too few lags may miss delayed coordination patterns; too many lags reduce statistical power and introduce noise. Optimal lag selection uses information criteria such as:

• Akaike Information Criterion (AIC)
• Bayesian Information Criterion (BIC) In collusion detection, domain knowledge about agent communication latency and decision cycles should inform the maximum lag window tested.

Granger causality has important limitations for detecting agent collusion:

• Linearity assumption: It captures only linear predictive relationships, missing complex nonlinear coordination strategies learned by deep reinforcement learning agents.
• No true causality: It measures predictive utility, not mechanistic causation. A third, unobserved coordinating agent could drive both observed agents' actions.
• Instantaneous effects: Granger causality cannot detect simultaneous coordination that occurs within the same time step.
• Omitted variable bias: Hidden communication channels or shared environmental triggers can confound results.

To overcome the linearity limitation, modern collusion detection systems employ nonlinear Granger causality methods:

• Kernel Granger causality: Maps time series into high-dimensional feature spaces to capture nonlinear dependencies
• Transfer entropy: An information-theoretic measure that quantifies directed information flow between agents without assuming linearity
• Convergent Cross Mapping (CCM): A technique from nonlinear dynamics that detects causality in deterministic, coupled systems These extensions are better suited for detecting emergent coordination in complex multi-agent reinforcement learning environments.

The systematic identification of unauthorized, covert coordination between autonomous agents. Granger causality provides the statistical foundation for collusion detection by testing whether one agent's historical actions contain predictive information about another's future behavior.

• Tests for temporal precedence in agent action sequences
• Distinguishes correlation from directional influence
• Flags statistically significant lead-lag relationships that warrant deeper investigation

A communication path enabling two agents to exchange information by manipulating shared system resources or timing mechanisms in violation of security policy. Granger causality analysis can detect covert channels by identifying statistically anomalous temporal dependencies between agents that should be operating independently.

• Detects timing-based information leakage
• Identifies resource-contention signaling patterns
• Reveals hidden communication through shared state variables

An indirect coordination mechanism where agents modify their shared environment to trigger specific actions from other agents, enabling complex emergent behavior without direct messaging. Granger causality helps distinguish stigmergic coordination from random environmental interaction by testing whether environmental modifications systematically precede and predict agent responses.

• Models environment-mediated agent influence
• Detects implicit signaling through shared state
• Differentiates intentional coordination from coincidental interaction

A state in MARL systems where independently trained agents learn to cooperate on a joint policy detrimental to overall system objectives, often by exploiting reward function flaws. Granger causality analysis of agent action trajectories can reveal emergent collusive policies by identifying mutual predictive relationships between agents' decision sequences.

• Detects emergent cooperative strategies not explicitly programmed
• Identifies reward-hacking coalitions
• Reveals implicit signaling through action selection patterns

The application of GNNs to learn normal interaction patterns in agent network topologies and identify anomalous nodes or edges indicating collusion. Granger causality features serve as edge weights or temporal attributes in GNN models, enriching the graph representation with statistically validated directional influence metrics.

• Encodes Granger-causal relationships as graph edges
• Combines structural and temporal anomaly signals
• Enables real-time collusion scoring on dynamic agent graphs

The technique of identifying a specific agent instance by analyzing unique statistical patterns in its decision-making, output distribution, or response latency. Granger causality testing can be incorporated into fingerprinting by establishing the distinctive temporal influence signature an agent exerts on its environment and peers.

• Profiles agent-specific causal influence patterns
• Tracks behavioral consistency across sessions
• Detects impersonation when temporal signatures deviate