When an artist's work is scraped into a generative model's training set, their only lever of control today is an honor-based exclusion file that most of them have never heard of: 59% of surveyed professional artists were unfamiliar with robots.txt, and 77% did not trust AI companies to respect it (Somesite I Used To Crawl: Awareness, Agency and Efficacy in Protecting Content Creators From AI Crawlers). Decentralized consent registries such as DECORAIT (DECORAIT - DECentralized Opt-in/out Registry for AI Training) were built to fix this, but they store licensing decisions on an append-only ledger: every time a creator changes a license, revokes consent, or reallocates royalties, the system must append a new transaction for every affected asset. A chameleon hash — a cryptographic hash with a secret trapdoor that lets a key holder swap the underlying message while keeping the hash output identical — offers exactly the missing primitive. To anyone without the trapdoor it behaves like an ordinary collision-resistant hash, so the ledger's audit structure stays intact; to the trapdoor holder, the record can be rewritten in place.
This proposal puts the trapdoor in the creator's hands. Existing redactable-ledger systems for AI — BlockFUL and its variants, and the "Proof of Federated Unlearning" protocol — all place the trapdoor with a consensus committee or platform operator (BlockFUL: Enabling Unlearning in Blockchained Federated Learning, Blockchain-enabled Trustworthy Federated Unlearning), which is precisely the agency relationship creators say they do not want. The core claim is that identity-based, key-exposure-free chameleon hashing (Identity-Based Chameleon Hash Scheme Without Key Exposure, Key-Exposure Free Chameleon Hashing and Signatures Based on Discrete Logarithm Systems) can give each creator a per-identity trapdoor over their own registry entries — supporting millions of independent rightsholders rewriting their own records without exposing anyone else's key, and with black-box traceability so edits remain attributable to the identity that made them.
Background and Gap
Three lines of work motivate the design. First, redactable blockchains established that chameleon hashes can reconcile ledger immutability with the need to rewrite history (Redactable Blockchain - or - Rewriting History in Bitcoin and Friends); later work added accountability through black-box traceability of trapdoor holders (Policy-Based Chameleon Hash with Black-Box Traceability for Redactable Blockchain in IoT, PIRB: Privacy-Preserving Identity-Based Redactable Blockchains with Accountability). None of this touches training-data consent. Second, creator-agency systems built the registry side: DECORAIT pairs an Ethereum ledger with C2PA manifests and ResNet-50 perceptual fingerprints, achieving 100% match accuracy on unmodified images and 91.2% under heavy distortion, with 4-millisecond queries over one million LAION images (DECORAIT - DECentralized Opt-in/out Registry for AI Training); Content ARCs extends this into a rights-and-compensation architecture but identifies metadata stripping and attribution scale as unsolved, and implements no cryptographic mutability mechanism (Content ARCs: Decentralized Content Rights in the Age of Generative AI). Third, the two existing chameleon-plus-AI systems — CHIP, which uses chameleon hashes for model-ownership passports (CHIP: Chameleon Hash-based Irreversible Passport for Robust Deep Model Ownership Verification and Active Usage Control), and BlockFUL, which uses them for committee-executed unlearning in federated learning (BlockFUL: Enabling Unlearning in Blockchained Federated Learning) — both place the trapdoor with the model owner or a committee, and neither addresses scraped-data contributors at all. The open slot is a consent ledger where the mutability key is held per-creator, and where edits are simultaneously invisible to the hash chain and attributable to their author.
Core Claim
A consent registry for AI training data can be built in which (a) each creator holds a per-identity trapdoor derived via identity-based key-exposure-free chameleon hashing, letting them rewrite their own license entries in place without a committee, without exposing master or sibling keys, and without invalidating any downstream hash pointer; (b) perceptual fingerprints serve as stable lookup keys so that rewritten licenses remain findable even when platform metadata is stripped; and (c) every rewrite is simultaneously digest-preserving and traceable, so an auditor can prove who changed what even though the block hash never moved.
Research Program
The system, which we call a Chameleon Consent Ledger, has three components.
The first component is the mutable registry entry. Each entry pairs a perceptual fingerprint (the stable key, following DECORAIT's ResNet-50 contrastive features) with a license object (the mutable value: opt-in status, permitted uses, royalty splits). The entry is committed with an identity-based chameleon hash, where a label — the creator's verified identity — parameterizes the trapdoor derivation. Concretely, in a discrete-log family scheme the public hash of a license $M$ with randomizer $R$ under identity label $\mathsf{id}$ takes the form $H = g^{M} y^{R} \bmod p$, and the holder of the per-identity trapdoor $x_{\mathsf{id}}$ can find $R'$ with $M + x_{\mathsf{id}}R \equiv M' + x_{\mathsf{id}}R' \pmod{p-1}$ for any new license $M'$, so the digest is unchanged (Key-Exposure Free Chameleon Hashing and Signatures Based on Discrete Logarithm Systems, Identity-Based Chameleon Hash Scheme Without Key Exposure). The critical property is key-exposure freeness: in naive schemes, an observer who sees both $(M,R)$ and $(M',R')$ can recover the trapdoor as $x = (M - M')/(R' - R)$, so a single edit leaks the long-term key. Identity-based granular trapdoors break that algebraic link, which is what makes per-creator mutability survivable at scale — one artist exercising a revocation cannot compromise the ledger or any other artist.
The second component handles the auditability paradox. A chameleon hash guarantees only syntactic equivalence: an outsider checking the digest cannot tell an edited entry from an untouched one, which is unacceptable when the whole point is proving that a consent change happened. We therefore adopt traceable chameleon hashing, in which a tracing algorithm can attribute a collision to the identity whose trapdoor produced it (Policy-Based Chameleon Hash with Black-Box Traceability for Redactable Blockchain in IoT, PIRB: Privacy-Preserving Identity-Based Redactable Blockchains with Accountability), and pair the live redactable chain with an immutable archive chain that logs redaction events — the dual-chain pattern proven out in federated unlearning by BlockFUL: Enabling Unlearning in Blockchained Federated Learning, which we reuse but invert the trust model: the trapdoor is the creator's, not a committee's.
The third component is the bridge to actual model behavior. A registry entry flip from opt-in to opt-out should trigger a recorded obligation: the trainer publishes a commitment to the post-unlearning model weights on the archive chain, referencing the redaction event, and the unlearning itself is executed with an existing method — ReTrack, which redirects the denoising trajectory of a diffusion model toward nearest retained neighbors, reporting an SSCD similarity of 0.37 with FID 8.45 on CIFAR-10 and SSCD 0.41 with FID 21.26 on CelebA-HQ (ReTrack: Data Unlearning in Diffusion Models through Redirecting the Denoising Trajectory). Compliance is then audited not by trusting the weight commitment but by running the operational copyright-evidence games of GenAI Copyright Evidence with Operational Meaning, which formalize "Access" and "Similarity" as membership-inference and reconstruction games with $(\alpha,\beta)$- and $(\eta,\gamma)$-evidence metrics. We are deliberate about scope here: this bridge provides tamper-evident records plus statistical evidence of compliance, not a cryptographic proof that the weights truly forgot — that stronger problem (noted as open in ReTrack and Blockchain-enabled Trustworthy Federated Unlearning) remains open, and this proposal does not claim to close it.
Proposed Experiments
The first experiment measures ledger economics on the same testbed DECORAIT used: the one-million-image LAION subset. We re-implement DECORAIT's E-FOF pipeline (off-chain retrieval over ResNet-50 fingerprints, on-chain event logs) exactly, then swap only the ledger layer for the Chameleon Consent Ledger, holding hardware, fingerprint code, and query routing fixed. The decisive metric is cumulative gas and wall-clock time under dynamic update workloads: initial ingestion, single license edits, and bulk revocation waves affecting 1%, 10%, and 50% of entries. DECORAIT's reference point is approximately 0.2M gas per ingestion with 4-millisecond queries and 100%/91.2% fingerprint accuracy; our system must match the accuracy and query latency while cutting cumulative update gas by a large factor, since an in-place chameleon rewrite touches one state cell whereas an append-only revocation writes a fresh event per asset.
The second experiment is a security and accountability evaluation. We test key-exposure resilience by having many creators perform collisions and then attacking the ledger: an adversary observing full before/after entry pairs attempts to recover the master key and sibling creators' trapdoors, and we report success rate (which should be negligible) alongside the cost of the tracing algorithm — proof size and verification time for attributing an anonymous edit to its identity — compared against a non-traceable chameleon baseline. This isolates whether the traceability layer's overhead, which the literature flags as substantial, is tolerable at registry scale.
The third experiment is the end-to-end revocation pipeline on CIFAR-10 and CelebA-HQ diffusion models, following the setups of ReTrack and GenAI Copyright Evidence with Operational Meaning. A creator flips their entry to opt-out; the trainer runs ReTrack and publishes the new weight commitment; the auditor runs the Access and Similarity evidence games. The key conditions are honest compliance (the pipeline should complete with acceptable overhead, and post-unlearning evidence should approach that of a model never trained on the revoked data) and spoofed compliance, where the trainer publishes unchanged weights — the evidence games should flag this at high detection rate. We also stress-test throughput under bursty revocation demand, mirroring BlockFUL's evaluation, which stayed stable up to 8,000 unlearning requests on its redactable live chain while sequential append-only pipelines degraded (Decentralized Federated Unlearning on Blockchain); our question is whether per-creator trapdoors, which remove the committee bottleneck, sustain the same throughput.
Baselines
The naive reference point is honor-based web exclusion — robots.txt and CDN bot blocking — which is the status quo for most creators. The empirical picture is discouraging: only about 14% of top-10k sites actively block AI crawlers, Cloudflare's blocking feature is used on 5.7% of eligible hosted sites, and AI assistant crawlers frequently ignore directives altogether (Somesite I Used To Crawl: Awareness, Agency and Efficacy in Protecting Content Creators From AI Crawlers). Any ledger-based system clears this bar trivially on enforcement; the real question is adoption friction, which we measure as creator-side setup cost (identity verification and trapdoor issuance time).
The strongest competitor is DECORAIT's E-FOF registry (DECORAIT - DECentralized Opt-in/out Registry for AI Training), which is the only existing system evaluated on the one-million-image LAION testbed and the closest match on the registry question. It reports 100% fingerprint accuracy on non-augmented images, 91.2% under heavy distortion, 4-millisecond queries, and roughly 0.2M gas per ingestion. We will run it unchanged on the same image set, same fingerprints, same retrieval index, and same hardware, with an identical tuning budget for the sharding parameters, so the only difference is the ledger layer. Beating it means three things together: fingerprint accuracy and query latency within a small factor of its reported numbers; cumulative gas under bulk license updates lower by at least an order of magnitude; and a capability it structurally lacks — creator-unilateral in-place updates with traceability. If the gas advantage is small at realistic update rates, the chameleon machinery is not earning its complexity and we will say so.
The key ablation is a committee-trapdoor variant: the identical Chameleon Consent Ledger but with a single shared trapdoor held by a designated committee, in the style of BlockFUL: Enabling Unlearning in Blockchained Federated Learning. This isolates the creator-agency ingredient — everything else (fingerprints, dual chain, traceability) is held fixed. We expect the committee variant to be marginally cheaper per edit (no per-identity derivations) but to require MPC-style key ceremonies for every update, adding latency; BlockFUL's own numbers show parallel redactions completing in 1.3–2.4 seconds versus 30–90 seconds for sequential retraining-based updates, which calibrates the latency budget a per-creator design must stay within. On the model side, ReTrack is both the execution engine and the unlearning baseline (SSCD 0.37, FID 8.45 on CIFAR-10; SSCD 0.41, FID 21.26 on CelebA-HQ, per ReTrack); we do not aim to beat its unlearning quality, only to show that wrapping it in the ledger-and-audit pipeline adds negligible latency and enables spoof detection it cannot provide alone.
Why not just…?
Why not just append revocation events to an ordinary smart contract? This is the most serious alternative, and for a creator with ten assets updating a license once, it is probably good enough — appending is simple and auditable by construction. The failure mode is scale and churn: at 0.2M gas per on-chain write even after DECORAIT's optimizations (DECORAIT - DECentralized Opt-in/out Registry for AI Training), a corpus of millions of assets with recurring license changes and royalty reallocations accumulates unbounded ledger state, and federated unlearning systems hit exactly this wall — sequential append pipelines saw throughput plummet under high request volume, forcing a move to redactable live chains (Decentralized Federated Unlearning on Blockchain). Our first experiment settles this empirically: we will report the gas ratio between append-only and in-place updates across update-rate scenarios, and if the ratio is under roughly three at plausible churn rates, the chameleon layer is not worth its overhead and the appendix-only design should win.
Why not just filter opted-out content before training, as The Stack does? The Stack shows pre-training exclusion works and does not hurt utility — 12.89% pass@1 on HumanEval with a 350M model, slightly above CodeGen-350M's 12.76% (The Stack: 3 TB of permissively licensed source code). We adopt this as the intake layer. But pre-filtering cannot address the central case this proposal targets: content already ingested, or consent revoked after training began. A static list offers no mechanism for post-hoc revocation, and metadata stripping severs the link between assets and their declarations — the persistence paradox documented in Content ARCs: Decentralized Content Rights in the Age of Generative AI. Our design keeps the fingerprint as the stable key precisely so revocation survives metadata loss.
Why not just train with differential privacy and skip the consent bookkeeping? Differential privacy gives a per-sample guarantee without any registry. The evidence against it as a wholesale substitute is that DP-Adam significantly degrades image generation quality in the CIFAR-10 diffusion evaluations of GenAI Copyright Evidence with Operational Meaning — the file reports no numeric value for the quality gap found on Lacuna, only the qualitative finding — and DP also does nothing for attribution or compensation, which are first-class requirements in the creator-agency literature (Computational Copyright: Towards A Royalty Model for AI Music Generation Platforms). That said, DP remains a reasonable complement for the access side, and our audit experiment can include a DP-trained model as an additional condition.
Why not just run ReTrack when an artist asks, with no ledger at all? If creators trusted trainers, this would suffice — ReTrack already achieves strong unlearning (0% frequency of unlearned images at FID 2.09 in the MNIST T-shirt setting, per ReTrack). But the trust deficit is the motivating condition: 77% of surveyed artists distrust AI companies to honor even voluntary web standards (Somesite I Used To Crawl: Awareness, Agency and Efficacy in Protecting Content Creators From AI Crawlers), and a trainer could claim to have unlearned while quietly retaining the weights. The honest caveat cuts the other way too: our ledger records obligations and commitments, and the audit layer provides statistical evidence via membership-inference and reconstruction games — it is not a cryptographic proof of unlearning, which no one has yet achieved for centralized generative models. The spoofed-compliance experiment is designed to quantify exactly how much detection power that statistical layer buys.
Why not just a trusted central registry with signed updates? If a single escrow party is acceptable, a database with signed mutations is cheaper and simpler than anything here, and we should admit that a cooperative single-jurisdiction setting does not need chameleon hashing. The target regime is zero-trust: the creator-agency literature converged on decentralized registries precisely because platform-custodied rights records reproduce the power asymmetry creators are protesting (Content ARCs: Decentralized Content Rights in the Age of Generative AI, DECORAIT - DECentralized Opt-in/out Registry for AI Training). Whether that convergence is right is a governance question our experiments cannot settle; we can only show that the decentralized, creator-held design is technically and economically feasible.
Decisive Checks
The proposal succeeds if three numbers come out favorably: cumulative gas for bulk license updates drops by at least an order of magnitude against append-only E-FOF at matched fingerprint accuracy and query latency; key-recovery attacks against the multi-creator ledger fail; and spoofed-compliance detection via the Access and Similarity evidence games is high while honest compliance incurs end-to-end overhead small relative to the unlearning run itself. It fails — and should be reported as failing — if the traceability overhead erases the gas advantage, if per-identity trapdoor issuance cannot be done at creator-signup cost, or if the evidence games cannot distinguish an honestly unlearned model from a spoofed one at useful rates. Each of these is measurable within the experiment plan above, on benchmarks with reported reference numbers.
References
- Somesite I Used To Crawl: Awareness, Agency and Efficacy in Protecting Content Creators From AI Crawlers
- DECORAIT - DECentralized Opt-in/out Registry for AI Training
- BlockFUL: Enabling Unlearning in Blockchained Federated Learning
- Blockchain-enabled Trustworthy Federated Unlearning
- Identity-Based Chameleon Hash Scheme without Key Exposure
- Key-Exposure Free Chameleon Hashing and Signatures Based on Discrete Logarithm Systems
- Redactable Blockchain - or - Rewriting History in Bitcoin and Friends
- Policy-Based Chameleon Hash with Black-Box Traceability for Redactable Blockchain in IoT
- PIRB: Privacy-Preserving Identity-Based Redactable Blockchains with Accountability
- Content ARCs: Decentralized Content Rights in the Age of Generative AI
- CHIP: Chameleon Hash-based Irreversible Passport for Robust Deep Model Ownership Verification and Active Usage Control
- ReTrack: Data Unlearning in Diffusion Models through Redirecting the Denoising Trajectory
- GenAI Copyright Evidence with Operational Meaning
- Decentralized Federated Unlearning on Blockchain
- The Stack: 3 TB of permissively licensed source code
- Computational Copyright: Towards A Royalty Model for AI Music Generation Platforms