Skip to content
Open access

Inter-xApp Policy Conflicts in the O-RAN Near-RT RIC: Threat Modeling, Attack Characterization, and Lightweight Conflict-Aware Mitigation

2026 · IEEE Access · Vol 14, pp. 151261-151273 · 0 citations · 17 references
Computer Science

Abstract

With the advent of the disaggregated, multi-vendor design of the Open Radio Access Network (O-RAN), third-party applications (xApps) can drive closed-loop control through the near-real-time RAN Intelligent Controller (near-RT RIC). In contrast to benefits alone, this openness creates a distinctively O-RAN threat: a malicious or misconfigured xApp can issue control actions that conflict with those of legitimate xApps over the same shared resource, without violating any cryptographic boundary. The O-RAN Alliance defines a Conflict Mitigation (CM) function for exactly this reason, yet inter-xApp policy conflicts remain mostly discussed conceptually rather than characterized and defended quantitatively. This paper studies the problem end to end on the shared mobility/handover resource. We formalize the threat using the STRIDE taxonomy and the O-RAN Alliance classification of direct, indirect, and implicit conflicts, highlighting a stealthy “solo-injection” channel that produces no direct multi-target conflict and thus evades naive signature detection. We then characterize a parameterized policy-conflict attack and compare, under an identical attack and metric set, three lightweight conflict-aware arbiters: a priority-based arbiter representative of existing CM frameworks, a deterministic, rule-based arbiter aligned with O-RAN guidance, and a shallow decision-tree arbiter that acts only on behavioral features and never on xApp identity. The unmitigated attack raises handover ping-pong from 0 to 23.3 events per UE per 100 slots and degrades service-level agreement (SLA) satisfaction from 0.963 to 0.776. The rule-based arbiter eliminates ping-pong but over-suppresses legitimate mobility, whereas the decision-tree arbiter intervenes selectively and attains the highest SLA (0.822) at $63{\,}\mu \mathrm {s}$ per prediction, well within the near-RT budget. We further present the defense’s sensitivity to scenario parameters and its failure modes. The machine-learning advantage is statistically significant and is preserved as the scenario grows to 30 UEs and 8 cells, and the arbiter tolerates 40% symmetric label noise with no loss of performance. It disappears, however, once the attack saturates the control loop, where the gate degenerates into the rule-based arbiter, and it inverts against an adversary that merely knows which behavioral quantities are observed: sampling its forged confidence from the benign distribution reduces detection recall from 0.9997 to 0.435 and leaves service quality below the undefended baseline. No single arbitration philosophy dominates, which argues for a hybrid design.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.