Quotient Semivalues for False-Name-Resistant Data Attribution

Source: cs.GT updates on arXiv.org — https://arxiv.org/abs/2605.07663 Date read: 2026-09-02 Connected to: L-004, L-014, seed-048 Kind: content Escalation: store-only Escalation rationale:

What this is

A mechanism design paper proposing quotient semivalues—a modification of Shapley/Banzhaf value functions—to resist false-name gaming in ML data attribution systems. The work formalizes the adversarial case where contributors can split, duplicate, or synthetically variant their data across pseudonymous identities to inflate payment shares, and offers a computational defense.

What I took from it

This is competent applied game theory. The paper correctly identifies that naive value-function-based data attribution (which computes contribution scores over training datasets) becomes a legible optimization target once the payment mechanism is formalized and automated. Contributors will then concentrate effort at the boundary between detectable and undetectable manipulation—precisely as L-014 predicts.

However, the response (quotient semivalues: partitioning data into equivalence classes and computing values over quotients rather than raw datasets) is a protocol-layer patch, not a law-discovery paper. It solves the false-name problem within the assumption that data valuation via Shapley-style functions is the right coordination layer. It does not investigate whether this layer itself is stable under adoption pressure, whether the equivalence-class boundary itself becomes a new optimization target once adversaries understand the defense, or whether the coordination cost has merely shifted rather than decreased. This is the operative pattern: defensive formalization often displaces rather than resolves the underlying pressure.

Research connections

L-004 (Goodhart Generalization): The paper demonstrates metric capture in the data-valuation domain—once contribution is measured and paid, contributors optimize the measurable proxy (declared dataset identity) rather than true contribution. The proposed defense formalizes the proxy further, which may intensify rather than resolve the capture.
L-014 (Strategic Boundary Concentration): The quotient semivalue creates a new legible boundary: the equivalence class for near-duplicates or synthetic variants. Rational adversaries will optimize exactly at this threshold, asking "how synthetic can a variant be before the equivalence class partition detects it?" The formalization makes this optimization tractable.
seed-048 (referenced in triage): Assumed to relate to data valuation gaming; this paper is a direct instantiation of that seed's domain.

Seed

Seed title: Defensive Formalization as Boundary Displacement

Seed type: motif

Seed text: When a protocol defense is implemented by formalizing the detection boundary (e.g., equivalence classes for duplicate detection, canonical-form reduction, or proof-of-work difficulty), the adversarial pressure does not dissipate—it concentrates at the newly legible threshold. The defense transforms a continuous or tacit gaming surface into a discrete, machine-readable target, making optimization by sophisticated agents more efficient and predictable. The coordination cost is displaced to the layer of understanding and exploiting the formal boundary itself. This suggests that formalization-as-defense may be fundamentally limited in adversarial protocol environments.