Privacy-Discounted Evidence: Letting Ignorant Teachers Abstain in Individualized PATE

When people who share data ask for different levels of privacy, a private learning system can honor those requests by giving each person's data a different weight. Data from people who want strong protection counts for less, so less noise is needed overall. Private Aggregation of Teacher Ensembles (PATE) is a common setting for this. It trains many "teacher" models on disjoint slices of private data and lets them vote, with noise, on labels for a public dataset that a "student" model then learns from. Its individualized version weights teacher votes by the privacy budget of the data behind them. That weighting contains a hidden failure. In practice, the people who demand strict privacy are often a distinct subpopulation. On queries from that subpopulation, the heavily weighted teachers have never seen anything similar, yet each still casts a full, confident vote and outvotes the few down-weighted teachers who actually know the answer.

We propose replacing each teacher's vote with an evidential output. Evidential deep learning trains a network to report how much evidence it has for each class, so it can say "I have no evidence here" instead of being forced to pick a class. Under this scheme an ignorant teacher contributes little or nothing, and down-weighted experts can still decide the label. The design rule that makes this safe is that privacy weights come only from users' publicly declared budgets. Evidence only shapes what a teacher says inside a fixed, bounded contribution. The privacy guarantee therefore holds no matter how badly calibrated the evidence is, which sidesteps the known unreliability of evidential uncertainty values.

Core Claim

Evidence in Subjective Logic, the belief calculus behind evidential deep learning, adds across sources, and discounting a source's evidence is a native operation. A heterogeneously weighted noisy vote is a discounted, noisy evidence sum. If each teacher outputs belief masses rather than a one-hot vote, the private aggregate is itself a noisy Dirichlet evidence count. Its sensitivity is the same as Individualized PATE's, because belief masses are nonnegative and sum to at most one. Its total mass also gives a principled answer-or-refuse test. We claim this changes accuracy for exactly the regime the heterogeneous-privacy literature flags as hard, namely when data and privacy demands are correlated. In that regime, evidential abstention should recover accuracy on the strict-privacy subpopulation at identical per-user privacy cost. When privacy demands are assigned independently of data, we expect little gain.

Background and Positioning

Heterogeneous privacy weighting supplies the mechanism. Heterogeneous Differential Privacy introduced per-item privacy levels. Optimal Data Acquisition with Privacy-Aware Agents shows that scaling a contribution by $w_i$ scales its sensitivity by $w_i$, and that optimal weights pool relaxed users and give strict users decaying individual weights. Two papers constrain the design. Personalized Privacy Amplification via Importance Sampling and Private Estimation When Data and Privacy Demands Are Correlated show that undoing down-weighting with $1/w_i$ factors destroys the privacy gain or blows up variance, so bias toward a prior is unavoidable. Evidential aggregation makes that bias explicit: down-weighted evidence shrinks the aggregate toward the flat Dirichlet prior instead of being reweighted. The second paper also motivates our central regime, because it treats correlated data and privacy demands as the hard case. Mean Estimation Under Heterogeneous Privacy: Some Privacy Can Be Free shows that optimal weighting saturates, so relaxing the lenient group's budget beyond a threshold buys nothing. This suggests that past some point, utility has to come from better use of strict-group information rather than from larger weights on the lenient group.

On the PATE side, two papers supply the method. Individualized PATE groups teachers by budget and weights their votes, but its teachers emit ordinary votes that cannot tell conflict apart from ignorance. Scalable Private Learning with PATE adds Gaussian noisy-max voting and a confident aggregator that refuses low-consensus queries, but only under a single global budget and over vote counts. We keep both mechanisms and change what each teacher contributes.

Evidential deep learning supplies the teacher output. Evidential Deep Learning to Quantify Classification Uncertainty maps nonnegative evidence $\alpha$ to belief masses $b_k=(\alpha_k-1)/S$ and vacuity $u=K/S$, where $S=\sum_k\alpha_k$ and $K$ is the number of classes. Two papers constrain how we use those quantities. Is Epistemic Uncertainty Faithfully Represented by Evidential Deep Learning Methods? proves that absolute evidential uncertainty is set by the regularization strength, drifts during training, and is reliable mainly for ranking. Beyond Magnitude shows that raw evidence magnitude conflates scale with confidence. We take these as a reason never to let evidence touch sensitivity, and we test robustness by sweeping the regularization strength.

The closest existing combinations are RESFL and BESplit. Both weight federated client updates by evidential uncertainty, and neither gives formal differential privacy. RESFL relies on adversarial attribute scrubbing, and BESplit explicitly leaves differential privacy to future work. Their design, in which weights are computed from private evidence, would leak if made formally private, because the weights are themselves functions of private data. Our proposal inverts the roles: public budgets set the weights, and evidence lives inside a bounded contribution.

Mechanism

Each teacher $t$ is trained with an evidential loss on its data partition. On a public query $x$ it outputs belief masses $b_t(x)\in\mathbb{R}^K_{\ge0}$ with $\|b_t(x)\|_1 = 1-u_t(x)\le 1$. The aggregator releases

$$A(x)=\sum_t w_t\, b_t(x)+\mathcal{N}(0,\sigma^2 I_K),$$

where $w_t\in[0,1]$ is the Individualized PATE weight fixed by the declared budgets of teacher $t$'s users, and $\sigma$ is the Gaussian noise scale. Each user's data enters one teacher. Any two belief vectors in this set differ by at most $\sqrt{2}$ in L2 norm, the same bound as for two one-hot votes. A user in teacher $t$ therefore faces sensitivity $\sqrt{2}\,w_t$. Per-query data-independent Rényi accounting is thus identical to weighted Individualized PATE, and this holds for any teacher-side function of private data, including poorly calibrated evidence.

A confident-aggregator variant uses the noisy aggregate's total mass. It answers only if the noisy top coordinate clears a threshold $T$, which is equivalent to requiring low vacuity $K/(K+\sum_k A_k)$ in the pooled Dirichlet. The data-dependent privacy analysis of Scalable Private Learning with PATE relies on vote-count gaps and does not transfer automatically to soft contributions. We will report data-independent bounds for every method and treat a data-dependent bound for soft evidence as a secondary theoretical goal.

Research Program

The work has three parts. The first is the mechanism and its accounting, including a careful proof that teacher-side calibration costs no privacy. The second is an evaluation protocol that, unlike prior PATE work, assigns privacy budgets in two regimes. In the independent regime, budgets are drawn at random, as in Individualized PATE. In the correlated regime, strict-privacy users form a coherent subpopulation. The third is a diagnostic study of why gains appear or not. It measures, on queries from the strict subpopulation, how often weighted lenient teachers outvote correct strict teachers, and how much belief mass ignorant teachers actually withhold.

Baselines

The naive reference point is standard PATE with Gaussian noisy-max voting at a uniform budget equal to the strictest user's, following Scalable Private Learning with PATE, which reports 98.5% MNIST accuracy at $\epsilon=1.97$ under a uniform budget. Beating it by a wide margin is expected and is not the target.

The strongest existing competitor is weighted Individualized PATE. It is purpose-built for heterogeneous budgets and reports 82.84% Adult accuracy against 79.85% for the non-individualized baseline. On MNIST it generates 890 labels against 257 for standard PATE at budgets $\log 2$ and $\log 8$. We will reproduce it on the same partitions, budgets, teacher architectures, and public query sets, with an equal hyperparameter search budget for the threshold and noise scale. All methods will use the same data-independent accounting, and the baseline's own reported accounting will be shown alongside. Beating it means higher student accuracy on the strict-privacy subpopulation at matched per-group $\epsilon$ in the correlated regime, without losing overall accuracy. A win only in the independent regime would not count as support for the claim.

Three further baselines isolate the key ingredient. The first gives teachers softmax-probability contributions that always sum to one, so they have soft votes but cannot abstain. The second has teachers abstain when maximum softmax probability falls below a threshold that is tuned per teacher with the same budget. The third keeps evidential teachers but uses uniform weights, which separates the effect of abstention from the effect of heterogeneous weighting. We will also sweep the evidential regularization strength to test whether gains survive the scale arbitrariness documented in Is Epistemic Uncertainty Faithfully Represented by Evidential Deep Learning Methods?.

Why not just…?

Why not just have softmax teachers abstain below a confidence threshold? This is the most serious threat, because a thresholded softmax teacher can also abstain and keeps the same sensitivity. The evidence from Evidential Deep Learning to Quantify Classification Uncertainty is that softmax stays confidently wrong on unfamiliar inputs such as rotated digits. In our setting, the lenient-subpopulation teachers would then still vote on strict-subpopulation queries, which is exactly the failure we target. That evidence comes from single models, not from ensembles under privacy. The thresholded-softmax baseline settles the question directly. If it matches the evidential teachers in the correlated regime, the contribution reduces to "abstention plus individualized weighting", and we will report it that way.

Why not just weight teachers or clients by their evidential uncertainty, as RESFL and BESplit do? Weights computed from private evidence are functions of private data. Using them inside the sensitivity calculation either breaks the guarantee or requires spending budget to privatize the weights. Those weights would also inherit the regularization-dependent scale described in Is Epistemic Uncertainty Faithfully Represented by Evidential Deep Learning Methods?. Our mechanism keeps evidence inside a norm ball whose size is fixed by public budgets, so these problems never arise.

Why not just use deep ensembles or Monte Carlo dropout per teacher? PATE already uses hundreds of teachers. Multiplying each by an ensemble or by many forward passes multiplies the cost of labeling every query. Single-pass evidential heads were measured at under 0.2% runtime overhead in BESplit. Within our setup, an ensemble teacher still has to turn its output into a bounded contribution, and nothing in the file shows it would separate ignorance from conflict better. We will run a small-scale Monte Carlo dropout variant on MNIST as a cost and accuracy check.

Why not just raise the weights of strict-privacy teachers? Doing so spends exactly the privacy those users refused to give. The saturation result in Mean Estimation Under Heterogeneous Privacy also suggests that reshuffling weights quickly runs out of gains. Our approach leaves every user's budget unchanged.

Proposed Experiments

We use MNIST, SVHN, and Adult with the budget groups of Individualized PATE, and FEMNIST for a natural writer-level subpopulation structure. In the correlated regime, strict budgets are assigned to a demographic group in Adult, to a class subset or style cluster in MNIST and SVHN, and to a set of writers in FEMNIST. We additionally inject notMNIST queries into the MNIST public set to test refusal on unanswerable queries.

The metrics are student accuracy overall and on the strict subpopulation, number of labels answered before the budget is exhausted, accuracy of the answered labels, and spent $\epsilon$ per group. We also report teacher test accuracy, because evidential training can cost some accuracy, which would offset the benefit.

The decisive check is the correlated-regime comparison against weighted Individualized PATE and against thresholded-softmax teachers at identical per-user $\epsilon$. The prediction is that evidential teachers reduce wrong outvoting on strict-subpopulation queries and raise strict-subpopulation student accuracy. In the independent regime the prediction is near-parity. The result should also hold across the regularization sweep. If the gain appears only at one regularization setting, or the thresholded-softmax baseline matches it, the claim fails as stated.