Drift as a Clock: Auditing When a Continually Trained Model Saw Its Data

Models that keep training after deployment do not sit still. Even when their accuracy stays flat, their internal representations keep wandering. Neuroscience calls this representational drift: the population of units that encodes a memory slowly changes while behavior stays the same. Data-auditing tools mostly ignore this. Membership tests, which ask whether a given example was in the training set, and differential audits are built and evaluated on a model frozen right after training. This proposal makes two claims. First, drift erodes the signals those audits rely on. Second, the same drift, if it proceeds at the right rate, carries a timestamp. Data learned at around the same time is stored in overlapping sets of units, so an auditor who plants a few marker documents at known times can read off when other data was ingested by comparing which units they engage.

The neuroscience result we borrow is from Drift of neural ensembles driven by slow fluctuations of intrinsic excitability. That work models a recurrent network in which each neuron's firing threshold wanders slowly. Each memory gets recruited into whichever neurons happen to be most excitable at that moment. A simple linear readout can then decode the day and the order of past memories from how much their neuron sets overlap. This works only when the threshold fluctuations are neither too slow nor too fast. We ask whether the same mechanism exists, or can be deliberately installed, in artificial networks trained on data streams. If it does, it addresses a gap that Datasheets for datasets acknowledges: static documentation cannot describe or verify evolving, streaming datasets.

Core Claim

The proposal has two linked parts. The first is a measurement claim about ordinary training. After a model has converged on a stream, keeping the training noise going causes two things to diverge. Loss-based membership signals for older data flatten toward chance. However, the unit-level "footprint" of that data continues to reflect when it was learned. The footprint is the pattern of which hidden units a batch most strongly engages.

The second is a design claim. Adding slow, per-unit gain fluctuations during training, an artificial analogue of excitability drift, sharpens this temporal tag. The decodability of ingestion time should then follow an inverted U as a function of the fluctuation timescale. That is the signature Drift of neural ensembles driven by slow fluctuations of intrinsic excitability reports in biology.

The new element is not detecting whether data was used. The new element is recovering ordinal ingestion time from a single current model, together with an explicit account of how the drift rate controls whether that information survives. The closest existing works are WARP: Weight-Space Analysis for Recovering Training Data Portfolios and Data Provenance via Differential Auditing. WARP recovers what proportion of the data came from each domain, using a single directed fine-tuning step. Differential auditing detects group membership in a frozen model. Neither models time, and neither has been tested under drift along flat, low-loss directions, where gradients are near zero.

Why drift should both destroy and preserve evidence

Two papers supply the mechanism, and they point in opposite directions for auditing. Representational drift as a result of implicit regularization shows that networks trained with noise past convergence move slowly along the set of equally good solutions toward flatter regions, sparsifying as they go. Stability through plasticity: Finding robust memories through representational drift shows the direction of that movement. Drift carries units away from "engaged" states, which respond in a near-linear, gradient-rich way, toward "disengaged" states that are inactive or saturated. The disengaged states are favored because there are many more of them and because they tolerate noise. It also shows that a separate allocation step is needed to re-engage units for new memories.

Differential auditing assumes the opposite. It assumes training data sits in sharp local curvature, where the model's output is sensitive to small input perturbations. If drift flattens that curvature, the differential signal should fade with time since ingestion. That is the tension the research file flags.

Allocation, however, implies a clock. Whatever units are engaged at time $t$ absorb the data seen at $t$, and the engaged pool rotates as drift proceeds. Co-ingested data therefore shares units, while data learned far apart in time does not. This is exactly the ensemble-overlap code in Purely STDP-based assembly dynamics: Stability, learning, overlaps, drift and aging, where a rise in learning rate moves networks from static to drifting assemblies. It also matches Dynamic and selective engrams emerge with memory consolidation, where neurons with higher initial firing rates remain in the memory ensemble longer.

The geometry of representational drift in natural and artificial neural networks constrains the design. Biological drift turns over many units and runs along high-variance directions, and node dropout reproduces this geometry in artificial networks. Our injected fluctuations should therefore be unit-wise, not isotropic noise added to the weights.

Research Program

The installed mechanism mirrors Delamare's moving threshold. Each hidden unit computes $$h_i = \phi\big(w_i^\top x + b_i - \theta_i(t)\big),$$ where $\phi$ is the nonlinearity, $w_i$ and $b_i$ are the unit's ordinary weights and bias, and $\theta_i(t)$ is a training-only offset. The offset follows a slow random walk that drifts back toward zero (an Ornstein–Uhlenbeck process), with timescale $\tau$ measured in optimizer steps and amplitude $\sigma$. At inference the offset is set to zero, so the deployed function is unchanged.

The footprint of a data set $S$ at layer $\ell$ is the vector $e_\ell(S)$ whose entry $i$ is the average, over examples in $S$, of the unit's activation times its loss gradient. This is a standard per-unit attribution score. Two sets are compared by the cosine overlap of their footprints.

The audit setting works as follows. In the spirit of Copyright Traps for Large Language Models, the auditor plants "anchor" batches into the stream at known times. The auditor then holds only the final model, plus candidate batches whose ingestion time is unknown. A small decoder maps each candidate's overlap profile against the anchors to an ingestion-time bin and an ordinal rank. This is the artificial counterpart of Delamare's "day decoder" and "ordinal time decoder."

The program has three stages. Stage one measures decay curves, meaning how each auditing signal weakens as a function of time since ingestion and of drift rate in unmodified training. Drift rate is varied through learning rate and gradient noise; the dropout rate controls the dropout-induced drift that The geometry of representational drift in natural and artificial neural networks identified. The signals are membership AUC, the differential-audit AUC, the WARP portfolio error, and footprint-based time decoding. Stage two adds the slow gain fluctuations and sweeps $\tau$ and $\sigma$, measuring the effect on both time decodability and task accuracy. Stage three tests whether the tag can be forged or erased by a provider who does not want to be audited. The attacks are short re-randomizing fine-tuning, weight averaging, and a deliberately frozen phase.

Baselines

The naive reference point is chance-level ordering. A decoder that ranks candidates randomly gets a Kendall rank correlation of zero, and its bin accuracy is one over the number of bins. A slightly stronger naive point uses the candidate's current loss alone.

The strongest existing competitor is a loss-gap forgetting curve. It applies a membership-inference score, either a loss ratio against a reference model or the Ratio MIA used in Copyright Traps for Large Language Models, and orders candidates by signal strength on the logic that older data is forgotten more. That paper reports a peak ROC AUC of 0.748 for membership with a 100-token sequence repeated 1,000 times, against about 0.52 without traps. We will reproduce membership AUC in our own small-model setting rather than reuse that number for timing. For temporal ordering itself, no reported result was found on Lacuna for any method. All decoders receive the same anchors, the same candidate batches, the same model access, and the same hyperparameter search budget, with decoders fit on held-out anchors. To count as beating this baseline, footprint decoding must achieve a clearly higher rank correlation in the post-convergence regime, where loss gaps have flattened.

Two further competitors cover the detection side. The first is differential auditing, Data Provenance via Differential Auditing, whose multiplicative variant reaches AUC around 0.80–0.81 on CIFAR-10, compared with 0.70–0.75 for shadow models. We extend it to timing by ranking candidates by differential magnitude. The second is WARP, WARP: Weight-Space Analysis for Recovering Training Data Portfolios, which reports portfolio MAE of 0.046 on BERT-base and 0.104 on GPT-2-small. We run WARP between consecutive snapshots in a privileged setting where the auditor has two checkpoints. This deliberately gives WARP more access than our method gets, and we will say so when reporting results.

The key ablation isolates timescale from noise. It compares three conditions at matched total noise power: slow gain fluctuations, fast fluctuations with the same $\sigma$ and very short $\tau$, and no fluctuation. If the fast condition decodes time as well as the slow one, the slow-excitability story is wrong.

Why not just…?

Why not just log timestamps or use zero-knowledge proofs of training? Trustless Audits without Revealing Data or Models can cryptographically prove that training ran on committed data. However, it requires the provider to generate proofs during training, and its authors note it is computationally difficult beyond small vision models. Our passive stage needs no provider cooperation; only the auditor's anchors must enter the stream. Our designed-fluctuation stage does require cooperation. There we concede that logging is simpler for an honest provider. The added value would be a tag embedded in the weights that survives the attacks in stage three. If those attacks erase it cheaply, that value disappears, and we will report this.

Why not just read ingestion order off the forgetting curve of membership scores? This is the real threat. If loss gaps fade monotonically with age, they already give a clock, and footprints add nothing. The two drift papers predict that after convergence, behavior-level signals stop changing while unit-level structure keeps moving. That predicts loss-based ordering degrades past convergence while footprint ordering does not. The decay-curve experiment tests this directly. If loss ordering matches footprint ordering at every drift rate, we drop the representational claim and report the forgetting-curve result as the finding.

Why not just freeze the model or add strong weight decay so there is no drift to worry about? Freezing preserves frozen-model audits but removes the clock. Stability through plasticity: Finding robust memories through representational drift argues it also locks units in fragile, densely engaged states. We include a frozen-phase condition and measure robustness to weight noise. This turns the trade-off into a number rather than an assertion.

Proposed Experiments

The first testbed is a ResNet trained on CIFAR-10, the differential-audit benchmark. It is trained to convergence and then continued for many "days" on a reshuffled stream with anchor batches inserted per day. The second is BERT-base continually fine-tuned on a stream over the ten Yahoo topic domains from the WARP suite, which lets us compare portfolio error and timing on the same model. The third is a small GPT-2-scale model with continued pretraining on successive Common Crawl snapshots of HuggingFaceFW/fineweb, whose 96 snapshots give a naturally time-ordered stream (The FineWeb Datasets). In the third testbed, trap-style anchor sequences are inserted at known snapshots. As a sanity check that the installed drift has the biological geometry, we compare its principal-component orientation and unit-turnover statistics against the Allen Brain Observatory analysis in The geometry of representational drift in natural and artificial neural networks.

The metrics are membership and differential-audit AUC as a function of age, Kendall rank correlation and bin accuracy for ingestion time, portfolio MAE, and clean task accuracy. Clean task accuracy must stay within a small tolerance of the no-fluctuation model.

Three outcomes decide the proposal. First, differential-audit AUC should fall with time since ingestion, and fall faster at higher drift rates. That would confirm the file's predicted tension between the differential-audit and drift-robustness papers. Second, in the post-convergence regime, footprint-based ordering should beat loss-forgetting ordering at matched access. Third, time decodability should peak at an intermediate fluctuation timescale and collapse in the fast-noise ablation at equal noise power. If the second outcome fails, the idea reduces to forgetting curves. If the third fails, the claim that a slow excitability process does the work fails.