Audited Delegation: Editorial-Board Structure for Verifying Cross-Organization Intrusion Alerts

When several organizations want to catch attacks that are spread thinly across all of them, each one eventually faces a question it cannot settle alone: "this looks odd here; is it part of something larger?" Today the usual answer is to ask peers directly, either by a vote or by querying them one at a time until confidence is high enough. Academic journals solved a structurally similar problem long ago. An editor-in-chief sends each submission to an associate editor who knows the area, that associate editor picks referees with the right expertise, and the editor-in-chief keeps the associate editors honest by occasionally checking their work. This proposal takes that structure literally as a protocol for verifying cross-organization intrusion alerts. It focuses on the part existing systems ignore: the middle tier is both what makes delegation efficient and the most dangerous place for a corrupted participant to sit.

The core claim is that a three-tier verification protocol has two properties that flat peer consultation lacks, and that both can be measured. Area-specialized intermediaries route alerts to the right experts, which cuts the number of peer consultations when no single node knows everyone's expertise. A randomized audit by the top tier, which sends a small fraction of alerts around the intermediary, bounds how much damage a corrupted intermediary can do before it is caught. If either property fails to show up against a strong flat baseline, the editorial analogy is decoration and should be dropped.

What exists and what is missing

The collaborative intrusion-detection literature supplies the mechanisms this proposal reuses. A Distributed Sequential Algorithm for Collaborative Intrusion Detection Networks treats alert verification as a sequential probability ratio test, meaning a node queries peers one by one, updates the odds that an intrusion is real after each answer, and stops once the odds cross a threshold set by the relative costs of false alarms, misses, and messages. Peer reliability in that paper is tracked with Beta-distribution reputation. Peers form a flat acquaintance list, and the paper assumes they are honest. Enhancing Challenge-based Collaborative Intrusion Detection Against Insider Attacks using Spatial Correlation hardens peer trust against "betrayal" insiders, who behave well until they have built up reputation, by combining synthetic challenge tests with passive monitoring of neighbors' message behavior. Its threat model concerns individual peers, not a supervisory layer. Decentralised Trustworthy Collaborative Intrusion Detection System for IoT comes closest to an editorial workflow, with contributors who submit rules, validators who vote on them, and regular nodes who deploy them. However, its validators check static signatures by consensus on a blockchain, at 3 to 5 seconds per rule, and it has neither area specialization nor any oversight of the validators themselves. Too Expensive to Attack: Enlarge the Attack Expense through Joint Defense at the Edge uses a central coordinator over edge agents and assumes the coordinator is fully trusted. These four papers supply the statistical machinery and the threat models. They are also where the gap shows: none of them models a corrupted intermediate tier.

On the organizational side, Multi-Agent Collaboration in Incident Response with Large Language Models finds that teams of language-model agents with a designated leader do somewhat better in a tabletop incident-response game. Centralized and hybrid teams reach 14 of 20 successes against 12 for heterogeneous decentralized teams. That gap is too small at 20 trials to support strong conclusions, and the setting has no real telemetry. Embodied LLM Agents Learn to Cooperate in Organized Teams shows that designated leadership cuts redundant communication among agents, with a 9.76% reduction in completion time. These papers motivate hierarchy but do not test it where it matters here. The principal-agent framing in Dilemmas and Governance Strategies in the Implementation of the Modern Apprenticeship System in Chinese Vocational Colleges: A Perspective from Principal-Agent Theory supplies the conceptual warning that delegated agents drift from the principal's goals as information asymmetry grows. The proposal turns that warning into an adversarial threat model.

The proposal deliberately does not use language-model agents as referees. Detection pipelines run at microsecond-to-second timescales: 173 µs per decision in ROSE-BOX: A Lightweight and Efficient Intrusion Detection Framework for Resource-Constrained IIoT Environments, and over $10^5$ requests per second in DBOS Network Sensing: A Web Services Approach to Collaborative Awareness. The editorial structure is implemented as a lightweight statistical protocol, not a conversation.

The protocol

There are three roles. Referees are the participating organizations' own detectors. Associate editors are participants elected to coordinate one attack area, such as volumetric floods, scanning, credential attacks, or web exploitation. Each associate editor maintains per-area Beta reputation for referees in its area, using challenge alerts with known ground truth in the style of the spatial-correlation paper. The editor-in-chief role, which can be held by a rotating participant or a jointly run service, triages each incoming alert to an area and sets stopping thresholds by Bayes risk. It also audits.

Verification inside an area follows the sequential test from the distributed sequential algorithm paper. The difference is that the associate editor orders referees by their area-specific reputation instead of using a flat acquaintance list. After $n$ referee verdicts $v_1,\dots,v_n$, the log-odds statistic is

$$ \Lambda_n=\sum_{i=1}^{n}\log\frac{P(v_i\mid \text{intrusion},\,e_i)}{P(v_i\mid \text{benign},\,e_i)}, $$

where $e_i$ is referee $i$'s estimated expertise in this area. Consultation stops when $\Lambda_n$ leaves the interval between the two Bayes-risk thresholds.

The audit is the part that is new. With probability $q$, the editor-in-chief withholds an alert from its associate editor and runs the same test itself on an independently sampled referee set. It then compares the outcome with what the associate editor would have reported, or later did report, on matched alerts. Suppose a corrupted associate editor tampers with a fraction $f$ of the $k$ alerts it handles, whether by routing them to weak or colluding referees, misreporting the verdict, or quietly dropping them. If audits are independent of the editor's choices, the chance of at least one audited tampering is

$$ P_{\text{catch}}=1-(1-qf)^{k}. $$

This follows directly from independent Bernoulli audits. It gives a design rule: the expected number of tampered alerts before detection is roughly $1/q$, independent of $f$. The audit rate therefore directly trades message overhead against how much harm a bad intermediary can do. Disagreements lower the associate editor's reputation. Fixed term limits with rotation address the betrayal pattern, where a node builds trust before turning.

Research Program

The work has three parts. The first is a simulator for cross-organization alerts built from public flow datasets, described below. In it, an attack campaign is split across $M$ simulated organizations so that each local share stays under that organization's own detection threshold. The second is the protocol plus its threat models: corrupted referees, a corrupted associate editor that suppresses alerts, a corrupted associate editor that colludes with referees, and a betrayal-timed associate editor. The third is an analysis of when the hierarchy pays off. We vary how much of the global expertise matrix any single node can observe, from full knowledge to each node knowing only its acquaintances. Our prediction is that routing through associate editors helps only when knowledge is local. If flat consultation matches it under full knowledge, that result is itself informative.

Baselines

The naive reference point is no collaboration at all. Each organization runs a strong local classifier, a ROSE-BOX-style feature-selected random forest, and alerts on its own. ROSE-BOX: A Lightweight and Efficient Intrusion Detection Framework for Resource-Constrained IIoT Environments reports 100% accuracy, precision, recall, and F1 on CSE-CIC-IDS2018 in the standard non-distributed split. That result shows the standard benchmark is saturated for local detection, so it cannot measure cross-system value. The local baseline is therefore run on our split campaigns, where we expect it to miss sub-threshold shares. If it does not miss them, our splits are not hard enough and must be made harder before anything else is claimed.

The strongest existing competitor is flat sequential consultation with Beta reputation from A Distributed Sequential Algorithm for Collaborative Intrusion Detection Networks. That paper reports about 3 consultations at peer expertise 0.7 and up to 50 at expertise 0.2 to reach the same confidence. These are the numbers our consultation counts will be compared against in its own simulated setting, which we reimplement with matching expertise levels. A stronger variant that we add ourselves orders flat peers by area-specific reputation. This is the key competitor, because it captures routing by expertise without any intermediate tier and without that tier's attack surface. A third baseline is challenge-plus-behavioral trust with weighted voting from Enhancing Challenge-based Collaborative Intrusion Detection Against Insider Attacks using Spatial Correlation, which reports detecting betrayal attacks 2 to 3 days faster than the earlier challenge-based model. We use that as the reference for time-to-detect a corrupted node.

Fairness rules apply to all methods. Every method sees the same alert streams, the same referee detectors, the same challenge-alert budget, and the same Bayes-risk cost settings. Thresholds and reputation priors get an equal tuning budget on a held-out week of simulated traffic.

Two ablations isolate the proposal's ingredients. One is hierarchy without audit ($q=0$), which shows how badly an unaudited middle tier fails under attack. The other is audit without area routing, which separates the robustness gain from the efficiency gain.

Beating the baselines means the following. Against honest participants, the full protocol must use fewer messages per verified alert than area-sorted flat consultation at equal detection rate and false-alarm rate, specifically in the local-knowledge regime. Under a corrupted associate editor, the number of tampered alerts before detection must stay near $1/q$, and the extra message cost must stay under the efficiency gain.

Why not just…?

Why not just match shared indicators cryptographically? Over-Threshold Multiparty Private Set Intersection for Collaborative Network Intrusion Detection finds IP addresses that appear at $t$ or more of 33 institutions in 170 seconds without revealing anything else. That is excellent for exact shared indicators. However, it answers "is this exact identifier seen elsewhere" and not "is this behavior part of a campaign." Our split campaigns include variants where attackers use different infrastructure at each organization. Running set intersection as an extra baseline on those variants shows directly where exact matching stops working. On campaigns that reuse identifiers, we expect set intersection to win, and we will report that.

Why not just federated learning? CGFL: A Robust Federated Learning Approach for Intrusion Detection Systems Based on Data Generation trains a shared intrusion model across clients with robust aggregation and reports an F1 of 0.98 on rare attack classes on NSL-KDD. Federated training improves every participant's local model, but the model still judges each local share alone. A share engineered to sit below threshold stays below threshold for a better local model too. Federated learning is therefore complementary: we will use a federated model as the referee detector in one condition, to check that the protocol's gain survives better referees.

Why not just flat consultation ordered by expertise? This is the real threat. If every node can see an accurate per-area expertise table, ordering peers by it should match the hierarchy's efficiency, and it has no middle tier to corrupt. The knowledge-locality sweep settles this. If flat ordering matches the protocol across the sweep, the efficiency claim is wrong and the remaining contribution is only the audit analysis. If the protocol wins only when knowledge is local, that defines the regime where editorial structure belongs.

Why not a language-model agent board, as in incident-response simulations? The supporting evidence in Multi-Agent Collaboration in Incident Response with Large Language Models is a 14-versus-12 difference over 20 runs, and agents need seconds per turn where verification needs milliseconds. Agents may suit a slow escalation channel for the few alerts the protocol cannot resolve, but they are not tested here as the core mechanism.

Proposed Experiments

The data come from CIC-IDS2017, CSE-CIC-IDS2018, and CIC-DDoS2019. Hosts are partitioned into 10 to 50 simulated organizations, and each attack campaign is sharded so that each organization's share falls below its local detector's threshold. Referee specialization is induced by training each organization's detector on a different subset of attack categories, which gives a known ground-truth expertise matrix. We also rerun the simulated network from the distributed sequential algorithm paper with expertise between 0.2 and 0.7, so that consultation counts can be compared directly with its reported figures.

The metrics are campaign-level detection rate and false-alarm rate, messages per verified alert, verification latency, number of tampered alerts before a corrupted associate editor is caught, and time to detect betrayal-timed nodes. There are three decisive checks. The first is whether the empirical catch curve under a corrupted associate editor follows $1-(1-qf)^k$; if it does not, audits are not independent of the attacker's behavior, and the protocol leaks information about which alerts are audited. The second is whether a crossover appears in the knowledge-locality sweep between the protocol and area-sorted flat consultation. The third is whether there is an audit rate $q$ at which the protocol is both cheaper than flat consultation and as robust as challenge-plus-behavioral trust. A negative answer to the second check would mean dropping the efficiency claim. A negative answer to the third would mean that editorial delegation costs more in security than it saves in messages.