When a learning system must keep trying options to discover which works best, and the feedback it learns from is private, it pays twice. It pays the ordinary cost of exploring, and it pays an extra cost because privacy noise blurs every observation, so it must explore longer. This extra cost, the regret-privacy gap, is largest early on, when few observations exist. That early phase is exactly where synthetic data is cheapest and most available, and synthetic data carries no privacy cost because it describes no real person. This proposal asks a precise question: how many synthetic observations are worth one real private observation, and how does that rate change as privacy tightens? We study it in a setting where all three pressures are real: hate speech moderation for low-resource Indian languages on energy-constrained devices.
The concrete system is a moderation node that, for each incoming post or voice clip, can afford to run only one of several detectors. The choices include a small text-only model, a monolingual model, and a heavier audio-plus-text model. The node discards the raw input afterwards. A community moderator later says whether the chosen detector's decision was right. Those labels are tied to users' content, so they are pooled across nodes only under differential privacy, a formal guarantee that the pooled output reveals almost nothing about any single user. Before deployment, each detector's accuracy can be estimated cheaply on LLM-generated functional test sets. Those estimates are free but biased, because aligned LLMs sanitize hate speech. Our claim is that a private bandit, an algorithm that repeatedly picks one option and learns only from that option's outcome, can use these synthetic estimates as a prior that cancels most of the privacy cost of exploration. The prior helps only while its bias stays below the gap between detectors. We build an algorithm that detects, cell by cell, when the bias crosses that line and stops trusting the prior.
Why this setting, and why it is a genuine bandit
The detector choice is a real decision with real stakes. In the DravidianLangTech@NAACL 2025 multimodal task, Overview of the Shared Task on Multimodal Hate Speech Detection in Dravidian languages reports that the top audio-plus-text systems reach macro F1 of 0.7511 in Malayalam and 0.7332 in Tamil. The best Telugu system reaches only 0.3817. Which detector is best therefore varies by language, and probably by script and by code-mixing within a language. On Marathi, L3Cube-MahaHate shows that a monolingual model beats multilingual ones (MahaBERT reaches 0.803 four-class accuracy). It also finds that cheap FastText-plus-CNN models remain competitive, so a cheaper detector is sometimes nearly as good.
Energy is what makes the feedback partial rather than complete. Batteryless and harvesting-powered devices run from a small capacitor that fills and empties with ambient power. Intermittent Inference with Nonuniformly Compressed Multi-Exit Neural Network for Energy Harvesting Powered Devices shows such devices must choose per input how much computation to spend, based on stored energy. That paper studies vision, where the device picks how deep to run one network. In our setting the device picks which detector to run at all. It cannot afford to run every detector, and it does not keep raw audio. So a moderator's label reveals only whether the chosen detector was right, not whether the others would have been. Without the energy limit and the discard policy, one label would grade every detector at once, and a bandit formulation would be unnecessary. We state this dependence openly, because it is the condition under which the proposal applies.
Privacy enters at aggregation. Privacy-Preserving Federated Learning for Hate Speech Detection motivates the concern. It trains hate speech classifiers on 1,543 Afrikaans and Russian sentences under DP-SGD (training with clipped, noise-added gradients). Accuracy falls from about 0.72 non-private to 0.602 at ε=5, and training collapses below roughly 20 sentences per client. That study is batch training, not sequential decision-making. It does show that, in low-resource hate speech, privacy noise overwhelms scarce real signal. We move this finding into the sequential setting, where the noise cost appears as excess regret, meaning cumulative wrong decisions relative to always using the best detector.
Core Claim
Recent work pins down the price of privacy in stochastic bandits. Optimal Regret of Bandits under Differential Privacy gives asymptotically optimal algorithms, DP-KLUCB and DP-IMED. Their per-arm cost is governed by a divergence $d_\epsilon$ that behaves like the usual KL divergence when privacy is loose and like a total-variation term scaled by ε when privacy is tight. The privacy part of regret therefore grows roughly like $\log T/(\epsilon\Delta_k)$ for an arm whose accuracy is $\Delta_k$ below the best. This matches the earlier lower bound of Shariff et al., which scales inversely with ε. Connecting Thompson Sampling and UCB offers DP-TS-UCB, which keeps total privacy loss constant while retaining optimal regret. These papers supply our method. All of them start each arm from zero knowledge.
We define a privacy-dependent exchange rate. Let $R_T(\epsilon; n_r, n_s)$ be the expected regret over $T$ rounds at privacy level ε, when each arm starts with $n_r$ real private observations and $n_s$ synthetic pseudo-observations. The exchange rate is
$$ \rho(\epsilon) = \frac{\partial R_T / \partial n_s}{\partial R_T / \partial n_r}, $$
which is the number of real private observations one synthetic observation is worth, at the margin. On the Equivalency, Substitutability, and Flexibility of Synthetic Data measured a static analogue for tracking, about 30 synthetic frames per real frame. The file notes that no exchange-rate formalism exists for text or for sequential learning.
We make two predictions. First, ρ(ε) rises as ε falls, because a real private observation becomes less informative under heavier noise while a synthetic one is unaffected. Second, ρ(ε) drops below zero once the prior bias $b_k$ for an arm exceeds about $\Delta_k/2$. At that point the prior reverses the ranking of two detectors, and every synthetic pseudo-count must be paid off with real evidence. The second prediction is where the hate speech literature constrains us. Delving into Qualitative Implications of Synthetic Data for Hate Speech Detection finds that LLM paraphrases lose their target group in over a third of cases and replace slurs with neutral identity terms. Synthetic test sets should therefore be most wrong exactly on neutral identity mentions and implicit hate. Synthetic priors will help in some cells and actively harm others, and a useful method must tell these cells apart.
Research Program
The algorithm, which we call test-then-trust, extends DP-KLUCB. Each pair of context cell and detector keeps two quantities. One is a public synthetic accuracy estimate with a pseudo-count $n_s$. The other is a private running sum of real outcomes, released in the geometric batches DP-KLUCB already uses. When a batch closes, the algorithm compares the private mean with the synthetic mean. The comparison uses a confidence bound that includes both sampling noise and the known Laplace privacy noise, built on the coupled concentration inequality from Optimal Regret of Bandits under Differential Privacy. If the two means are inconsistent, the cell's synthetic pseudo-counts are discarded permanently. Otherwise they keep contributing to the index the algorithm uses to pick a detector.
Because synthetic data and the test threshold are public, this adds no privacy cost beyond DP-KLUCB's. We will prove an upper bound of the form: regret equals the cold-start private regret, minus a saving in cells where the bias is small relative to the gap, plus a bounded detection cost in cells where it is large. That detection cost should be logarithmic and should not scale with $n_s$. We will also derive ρ(ε) in closed form for Bernoulli rewards. The proofs will be asymptotic, like the results they build on. We will report finite-horizon behavior empirically and not claim it theoretically.
Energy enters as an availability constraint. Each detector has an energy cost profiled on a reference device. A simulated capacitor, charged by a stochastic harvesting process, makes expensive detectors unavailable when stored energy is low. This is a "sleeping arms" condition, and we will not claim to solve the full energy-budget problem. A second energy effect concerns privacy accounting. A brownout can interrupt a geometric batch before it closes. Following the checkpointing logic of Lazarescu and Balsamo and Zhang et al., we will store the batch counter and partial sums in non-volatile memory, so that privacy accounting stays tied to observations actually made. The energy model is simulated. We do not claim transformer inference runs on an MSP432-class chip.
The synthetic priors come from two sources. For the Indian-language streams, we generate LLM paraphrase test sets per functional category, following the recipe of Casula et al. For Southeast Asian languages, the silver and gold sets of SEAHateCheck let us measure prior bias directly. Each detector's accuracy on silver tests minus its accuracy on gold tests, per category, gives a real estimate of $b_k$.
Baselines
The naive reference point is synthetic-only selection. For each context cell, we pick the detector with the highest synthetic accuracy and never explore. It uses no privacy budget, and its regret grows linearly wherever the synthetic ranking is wrong. It shows the cost of trusting generated data without checking it.
The strongest existing method is cold-start DP-KLUCB from Optimal Regret of Bandits under Differential Privacy, alongside DP-TS-UCB from Connecting Thompson Sampling and UCB. It is the strongest because it is asymptotically optimal among private algorithms that use no side information. Any gain we report must therefore come from the synthetic prior, not from a better private bandit. Its reported advantage is "up to 10 times lower regret" than DP-SE on synthetic Bernoulli bandits. No raw regret values for DP-KLUCB or DP-TS-UCB on hate speech streams were found on Lacuna. We run these baselines with the same streams, the same ε values, the same detector pool, the same energy traces, and the same tuning budget for batch parameters. Beating them means lower mean cumulative regret at matched ε, with a margin outside the across-seed confidence interval, at ε ≤ 1 where the privacy term dominates.
The key ablation is naive warm start. It is identical to our method except that synthetic pseudo-counts are never discarded. It isolates the bias test. Our method should match it where bias is small and beat it clearly where bias is large, especially in neutral identity mention cells.
A non-private KL-UCB with and without warm start gives the floor. The difference between it and its private counterpart measures the regret-privacy gap directly, so we can report what fraction of that gap the prior closes.
The detectors themselves are existing systems. They include MahaBERT and FastText-CNN for Marathi, as in L3Cube-MahaHate. For Dravidian languages they include XLM-RoBERTa text-only models and Whisper or Wav2Vec 2.0 late-fusion models, following the top shared-task teams. The detector-level accuracies cited above anchor how large the gaps $\Delta_k$ can be.
Why not just…?
Why not just pick the detector offline on the synthetic test suite and skip exploration? This is the synthetic-only baseline. It fails wherever generated data misranks detectors. The Robustness Paradox reported by Casula et al. gives a reason to expect misranking. Synthetic training lifts HateCheck macro F1 from .386 to .665, yet the same data erases targets and swaps slurs for neutral terms. A synthetic benchmark can therefore systematically prefer the wrong detector on identity-term cells. The SEAHateCheck silver-versus-gold bias measurements will show directly whether such misrankings occur. If they never do, this simple approach wins and our method is unnecessary. We treat that as a possible outcome, not a strawman.
Why not just spend a small privacy budget once, on a held-out gold set, and select the best detector with a one-shot private selection? This is the most serious alternative. With few detectors, few context cells, and stable accuracies, a single noisy selection on a few hundred labeled items may match any bandit. We run it at the same total ε and the same label count. We expect the bandit to win only when cells are numerous and labels per cell are few, which is exactly the low-resource regime, or when accuracies drift over time. If one-shot selection matches us across the grid of cell counts, we will report that and narrow the claim to the exchange-rate analysis.
Why not just use local privacy, with each moderator's label randomized before it leaves the device? This avoids trusting a server, but Basu et al. show local privacy multiplies regret by a factor scaling with $1/\epsilon^2$. Under that model the exchange rate should be even larger. We include a local-privacy variant to test this, since it would strengthen the case for synthetic priors.
Why not just translate posts to English and moderate centrally? That needs raw content to leave the device, which the discard-and-privacy setting rules out. It also loses the monolingual advantage documented for Marathi. We do not test this route, because it violates the deployment constraint the proposal assumes.
Proposed Experiments
We will build replay streams from three sources. The DravidianLangTech multimodal data provides Malayalam, Tamil, and Telugu clips, with cells defined by language and text-only versus speech availability. L3Cube-MahaHate provides 25,000 Marathi tweets, with cells defined by class-relevant topic and script. The Afrikaans/Russian set provides a very small, privacy-motivated stream. Each stream will be shuffled across 20 seeds. Detectors run on held-out folds, so each arm's outcome on each item is precomputed, but the bandit sees only the outcome of the arm it chose.
The first decisive check is the exchange-rate curve. We will estimate ρ(ε) at ε from 0.1 to 10 and expect it to increase monotonically as ε decreases. A flat curve would refute the central claim.
The second check maps the bias boundary. We will run a semi-synthetic Bernoulli experiment in which prior bias is injected in controlled amounts. Regret curves for naive warm start should cross cold start near $b_k \approx \Delta_k/2$, while test-then-trust should stay below both.
The third check looks at where the bias falls in real data. Using the SEAHateCheck silver/gold differences and our Indian-language paraphrase sets, we expect the bias test to discard priors mainly in neutral identity, counter-speech, and negation cells, as the sanitization finding predicts. A random pattern of discards would suggest the test is merely reacting to noise.
The fourth check tests the energy constraints. We will vary harvesting intensity so that the fusion detector is often unavailable, and confirm two things. First, the regret advantage persists under sleeping arms. Second, checkpointed batch accounting keeps the realized privacy loss at its nominal value, while an uncheckpointed version either over-counts or mis-pairs observations.
Our metrics are cumulative regret in misclassified items, the fraction of the private-minus-non-private regret gap closed, per-cell discard rates, and realized privacy loss.
References
- Overview of the Shared Task on Multimodal Hate Speech Detection in Dravidian languages: DravidianLangTech@NAACL 2025
- L3Cube-MahaHate: A Tweet-based Marathi Hate Speech Detection Dataset and BERT Models
- Intermittent Inference with Nonuniformly Compressed Multi-Exit Neural Network for Energy Harvesting Powered Devices
- Privacy-Preserving Federated Learning for Hate Speech Detection
- Optimal Regret of Bandits under Differential Privacy
- Differentially Private Contextual Linear Bandits
- Connecting Thompson Sampling and UCB: Towards More Efficient Trade-offs Between Privacy and Regret
- On the Equivalency, Substitutability, and Flexibility of Synthetic Data
- Delving into Qualitative Implications of Synthetic Data for Hate Speech Detection
- ARM mbed support for transient computing in energy harvesting IoT systems
- Low Overhead Online Data Flow Tracking for Intermittently Powered Non-Volatile FPGAs
- SEAHateCheck: Functional Tests for Detecting Hate Speech in Low-Resource Languages of Southeast Asia
- Differential Privacy for Multi-armed Bandits: What Is It and What Is Its Cost?