Colluding Adversaries in ML Pipelines

Overview

This series explains how colluding adversaries in ML pipelines turn one attack’s outcome into another attack’s prerequisite. It maps trust boundaries, roles, capabilities, and knowledge; examines train-test and test-time dependencies; evaluates amplification, reversal, and nonlinear case-study results; and develops edge-focused defenses, including the wider supply-chain and runtime context. The framework is conditional and correlation-based, not a universal risk score or causal guarantee.

Learning outcomes

Collusion requires one attack to supply a useful condition or artifact for another; shared vulnerabilities alone are insufficient.
Train-test paths depend on how poisoning changes tail behavior or distinguishability, not aggregate accuracy alone.
Test-time artifacts gain value from quality, overlap, similarity, access, timing, and query budget; more data can eventually add noise.
Case studies show amplification, reversal, and nonlinear effects, so dependency edges are hypotheses to test, not causal guarantees.
Defenses should interrupt high-value transfers and be reevaluated across accuracy, privacy, fairness, robustness, and cost.

Keywords

collusionMLpipelinespoisoningevasionprivacyinferenceextractionreconstructiondefenses

Chapters

Colluding Adversaries in ML Pipelines: The Blind Spot

Why attack outcomes become another attack’s prerequisites

13 min · Why attack outcomes become another attack’s prerequisites

13 min

Colluding Adversaries in ML Pipelines: Trust Boundaries

Map who can change, observe, or pass forward state

16 min · Map who can change, observe, or pass forward state

16 min

Colluding Adversaries in ML Pipelines: Poisoning at Deployment

How training changes downstream susceptibility

18 min · How training changes downstream susceptibility

18 min

Colluding Adversaries in ML Pipelines: Test-Time Collusion

How queries become reusable data, knowledge, or models

16 min · How queries become reusable data, knowledge, or models

16 min

Colluding Adversaries in ML Pipelines: Case Study Evidence

What baselines, reversals, and nonlinear results establish

16 min · What baselines, reversals, and nonlinear results establish

16 min

Colluding Adversaries in ML Pipelines: Partners and Chains

When stronger attackers create or remove an edge

15 min · When stronger attackers create or remove an edge

15 min

Colluding Adversaries in ML Pipelines: Edge Defenses

Protect transfers across data, models, queries, and actions

19 min · Protect transfers across data, models, queries, and actions

19 min

Preview every chapter · Unlock full text, audio, and video from chapter 4

Plans

Ch.1-3 FreeCh.4-7 Premium

FAQ

What is collusion in an ML pipeline?

Collusion is explicit coordination in which one adversary’s attack outcome improves another adversary’s attack by supplying a needed condition, artifact, or knowledge.

How can poisoning help a later attack?

Poisoning can help when it changes tail length, dataset distinguishability, or subgroup distinguishability in a direction that a deployment-time attack can exploit; the relationship remains conditional.

When is a collusion partner unnecessary?

A partner is unnecessary when the downstream attacker already has the needed model knowledge, data overlap, distribution knowledge, or other prerequisite directly.

How should defenders assess colluding attacks?

Defenders should map roles, artifacts, prerequisites, quality, timing, and access, then test isolated, combined, and adaptive paths against explicit baselines.