Probability Hacking and the Design of Trustworthy ML for Signal Processing in C-UAS: A Scenario Based Method

In order to counter the various threats manifested by Unmanned Aircraft Systems (UAS) adequately, specialized Counter Unmanned Aircraft Systems (C-UAS) are required. Enhancing C-UAS with Emerging and Disruptive Technologies (EDTs) such as Artificial Intelligence (AI) can lead to more effective countermeasures. In this paper a scenario-based method is applied to C-UAS augmented with Machine Learning (ML), a subset of AI, that can enhance signal processing capabilities. Via the scenarios-based method we frame in this paper probability hacking as a challenge and identify requirements which can be implemented in existing Rule of Law mechanisms to prevent probability hacking. These requirements strengthen the trustworthiness of the C-UAS, which feed into justified trust - a key to successful Human-Autonomy Teaming, in civil and military contexts. Index Terms: C-UAS, Scenario-based method, Emerging and Disruptive Technologies, Probability hacking, Trustworthiness.

💡 Research Summary

The paper addresses the emerging challenge of integrating machine‑learning (ML)‑based signal‑processing capabilities into Counter‑Unmanned Aircraft Systems (C‑UAS) while safeguarding the Rule of Law (RoL) that underpins societal trust. The authors argue that although Artificial Intelligence (AI) as an Emerging and Disruptive Technology (EDT) can dramatically improve detection, tracking, and classification of hostile drones, it also introduces a new class of legal‑ethical risk they term “probability hacking” (p‑hacking). In the context of ML, p‑hacking refers to the manipulation or selective reporting of performance metrics that leads to inflated accuracy claims, which can undermine justified trust in autonomous decision‑making, especially in high‑stakes military or civil security operations.

To systematically address this risk, the authors develop a four‑step scenario‑based method:

1. Capability Gap Analysis – They map the current C‑UAS kill‑chain (prevention, detection, tracking, classification/identification, designation of intent, engagement, forensics) and identify a concrete gap: radar detection of small UAS (sUAS) with low Radar Cross Section (RCS) in cluttered environments. Conventional radar struggles with low‑SNR returns, motivating the use of a Recurrent Neural Network (RNN) to learn temporal‑frequency patterns and improve detection rates.

2. Scenario Design – A realistic operational scenario is constructed that spans both civil security (e.g., protecting a large public event) and military contexts. The scenario explicitly incorporates four legal factors that determine legitimate use of force: legitimacy (self‑defence, UN mandate, aid), type of conflict (IHL), environment (urban, rural, etc.), and minimisation of impact on critical infrastructure. By embedding the ML system within this legally‑qualified context, the focus shifts from “is the use of force lawful?” to “does the ML component unintentionally erode the RoL?”

3. Requirement Identification – Drawing on responsible AI frameworks and ML‑ops (MLOps) best practices, the authors derive a set of interdisciplinary requirements:

   • Pre‑registration of research design (algorithm, data sources, hyper‑parameters, validation plan) to create a transparent audit trail.
   • Public disclosure of performance evaluation protocols and independent verification of results.
   • Explicit documentation of data provenance, bias mitigation, and error characteristics before deployment.
   • Legal‑by‑design clauses that tie technical specifications to statutory obligations (e.g., proportionality, distinction between civilian and military objects). These requirements are positioned as extensions to existing RoL mechanisms (policy revisions, procurement standards, and judicial oversight).

4. Validation, Verification, and Implementation – The paper outlines how the above requirements can be operationalised within R&D and procurement pipelines. For instance, a pre‑registered design is submitted to an independent oversight body; performance metrics are stored in immutable repositories; compliance checks are embedded in contract clauses; and post‑deployment monitoring includes both technical (drift detection) and legal (rule‑of‑law impact assessment) feedback loops.

A concrete “implicit p‑hacking” case study illustrates the danger. A start‑up wins a civil‑security tender by reporting high detection rates for both urban and rural settings. However, the RNN was trained exclusively on Fixed‑Wing (FW) UAS data, while the actual threat at a large urban event involves Rotary‑Wing (RW) sUAS operating in dense clutter. The hidden dataset bias leads to a dramatic drop in real‑world detection performance, exposing operators to unanticipated risk. The authors argue that without the pre‑registration, transparency, and legal‑by‑design safeguards they propose, such hidden biases remain undetected until catastrophic failure.

The paper concludes that the scenario‑based method provides a structured pathway to reconcile technological innovation with the RoL. By embedding legal‑ethical analysis early in the design cycle, C‑UAS equipped with trustworthy ML can achieve “justified trust”—the confidence required for effective Human‑Autonomy Teaming (HAT) in both civil and military domains. The authors call for institutional adoption of their interdisciplinary requirements, suggesting that only through combined technical transparency, legal oversight, and operational accountability can probability hacking be mitigated and trustworthy autonomous systems be realized.