When phones collaboratively train a speech recognizer through federated learning, each device keeps its audio and sends only model updates (gradients) to a server. Recent work shows this is not enough protection. An attacker who sees one update can run gradient inversion, which means searching for an input whose gradient matches the one observed. This recovers up to 140 seconds of audio, and the recovered speech is good enough that a standard speaker-verification model identifies who said it. Two standard fixes both have serious costs. Adding differential-privacy noise to the updates hurts accuracy unless tens of millions of users take part. Running a full voice-anonymization system on every phone before training is expensive and leaves residual identity behind.
We propose a simpler structural fix. Put a frozen, publicly trained encoder at the bottom of the on-device model, and train that encoder beforehand to discard speaker identity. Then federate only the layers above it. Every quantity the device sends is then a function of the encoder's speaker-scrubbed output, never of the raw waveform. So no attacker, however clever or however much it controls the server, can extract more identity from the updates than a probe could extract from that output. The question of how much voiceprint the gradients leak becomes a question you can measure once and offline: how much identity survives in the frozen representation. The research asks where to cut the network and how hard to scrub so that this identity bound is near chance while word error rate barely moves.
Core Claim
Let $x$ be a client's audio, $S$ its speaker identity, $z = E(x)$ the output of the frozen encoder $E$, and $G$ the update the client sends for the federated upper layers with parameters $\theta$ and transcript $y$. Because $G$ is computed from $(z, y, \theta)$ alone, any attacker output $A(G)$ satisfies
$$ I(A(G); S) \le I(z, y; S), $$
where $I$ is mutual information. This follows from the data-processing inequality: processing $z$ further cannot add information about $S$ that $z$ did not already carry. The bound still holds if the server picks $\theta$ maliciously, since $\theta$ cannot restore information $E$ removed.
In practice we replace mutual information with an operational stand-in: the equal error rate (EER, the operating point where false accepts equal false rejects, so 50% is chance) of the strongest speaker-verification probe we can train on $z$ sequences. The claim to test has three parts. The probe EER upper-bounds the privacy of every gradient-level attack. It can be driven near 50%. And the word-error cost of doing so is small, because deep speech representations already separate linguistic content from speaker traits. A precedent for the last point is On the Encoding of Gender in Transformer-based ASR Representations, which erased a linear gender subspace from Wav2Vec2 with WER moving only from 2.76% to 2.77%.
We state the scope plainly. This protects the voiceprint, not the words. The transcript content in $z$ is exactly what a speech recognizer must learn from, so inverted updates may still reveal what was said. We measure content leakage separately and treat it as a job for differential privacy layered on top, applied to a smaller and better-conditioned set of parameters.
How This Relates to Prior Work
Several papers motivate the threat. How to Recover Long Audio Sequences Through Gradient Inversion Attack With Dynamic Segment-based Reconstruction splits a long spectrogram into segments and inverts each against the shared gradient. It reports an average cosine similarity of 0.98 and 47.7 dB SNR on Speech Commands, and explicitly leaves defenses open. We take its attack as the primary adversary. Privacy Attacks for Automatic Speech Recognition Acoustic Models in A Federated Learning Framework does not invert anything. It feeds a 32-minute public dataset through the global and adapted models and verifies speakers from the activation differences, reaching 1–2% EER overall and 0.86% from the first hidden layer alone. That result tells us the lowest layers are where identity lives, which is why we freeze and scrub them rather than federate them.
Other work supplies the scrubbing methods. The speaker-subtraction and gradient-reversal training in EASY: Emotion-aware Speaker Anonymization via Factorized Distillation and the linear concept erasure used in the gender study are the tools we reuse, applied to the encoder's representation rather than to re-synthesized audio.
Another group constrains the design. Analysis of Speech Temporal Dynamics in the Context of Speaker Verification and Voice Anonymization shows that speaking rhythm alone re-identifies speakers even when timbre is masked. Scrubbing each frame independently is therefore insufficient, and our probe has to operate on whole sequences. Evaluating Voice Conversion-Based Privacy Protection against Informed Attackers shows that attackers who know the transform can invert it. We accordingly assume the attacker knows $E$ exactly. The bound does not depend on keeping $E$ secret.
The closest defenses are image-domain gradient defenses: decoy samples plus projection in Concealing Sensitive Samples against Gradient Leakage in Federated Learning, and client-private mixed-precision quantization in Mixed Quantization Enabled Federated Learning to Tackle Gradient Inversion Attacks. Both make inversion harder. Neither limits what a successful inversion could contain, and neither has been tested on speech.
In speech, General or Specific? Investigating Effective Privacy Protection in Federated Learning for Speech Emotion Recognition perturbs attribute embeddings and re-synthesizes audio on the device. This protects a pre-chosen demographic label but has not been tested against full reconstruction. We differ in removing identity inside the model rather than in the audio, and in targeting identity itself rather than one attribute.
Research Program
The first phase builds the encoders. We start from a pretrained self-supervised speech encoder. For cut depths ranging from shallow to deep, we train a scrubbing head on public data (LibriSpeech and VoxCeleb) to remove speaker identity from the output at that depth. We use two scrubbers. One is a linear eraser. The other is adversarial training with a gradient-reversal speaker classifier that sees whole sequences, so rhythm cues are penalized too. A connectionist temporal classification (CTC) loss on public transcripts keeps the representation useful for recognition.
The second phase certifies each encoder offline. We train the strongest sequence-level speaker probe we can, with ECAPA-style pooling over $z$, full enrollment data, and a tuning budget matched to the attacker's. Its EER is the claimed privacy bound for that encoder.
The third phase federates the upper layers on held-out speakers. We then attack the resulting updates three ways. The first attacker optimizes a waveform through the known frozen $E$ using segment-based gradient matching. The second is the indicator-dataset probe. The third is an active server that chooses $\theta$ to maximize leakage. We score all three by running ECAPA-TDNN verification and Whisper transcription on whatever they recover. The central empirical test is whether every attack's EER stays at or above the offline probe's EER. If an attack beats the bound, either the attacker is stronger than our probe or there is a leak we have not modeled, and either outcome is reportable.
Baselines
The naive reference point is ordinary federated fine-tuning of the whole model with no defense, attacked by the segment-based inversion. We expect it to reproduce the published attack strength of 0.98 cosine similarity and 47.7 dB SNR from the dynamic segment-based inversion paper, and near-perfect speaker verification on reconstructions. This run calibrates the attack pipeline.
The strongest formal competitor is user-level differential privacy with per-layer clipping, the LAMB optimizer, and a public seed model, following Federated Learning with Differential Privacy for End-to-End Speech Recognition. That paper reports an absolute WER increase of 1.3% to 4.6% at $\epsilon < 10$, but only with about 70 million simulated users. We will run its recipe at our cohort sizes of hundreds to thousands of clients, with the same seed model, data, and number of tuning trials as our method. We beat it if, at matched attack EER on reconstructions, our WER is lower. We also run the combination of our encoder with DP on the upper layers, since that is the configuration we would actually deploy.
The strongest speech-side competitor is on-device anonymization before local training, using a VoicePrivacy-style re-synthesis system. The file gives two reference points. Introducing the VoicePrivacy Initiative reports residual EER falling to 26–37% against attackers who also anonymize their enrollment data, with relative WER increases of 19–67%. EASY reports about 46.7% EER at 2.7% WER. We apply the anonymizer to client audio, federate the full model, attack the updates, and report on-device compute. We beat it if we match its attack EER at equal or lower WER and at a small fraction of its compute.
For gradient-obfuscation defenses, we port mixed-precision quantization and decoy projection to the speech model under the same tuning budget. There is no reported result on Lacuna for either method on speech or recognition tasks. Their image results, a PSNR drop to 8.1 on CIFAR-10 for quantization and to about 7.84 on MNIST for decoys, cannot be transferred directly.
The key ablation is freezing at the same depth without scrubbing. It isolates whether the scrubbing step matters or whether depth alone does the work.
As a secondary task, we train speech emotion recognition on CREMA-D and compare against Pro-Ind. Pro-Ind's reported number is about 66% accuracy, against below 55% for user-level DP.
Why not just…?
Why not just freeze the lower layers without scrubbing? This is the most serious threat to the proposal. If deep representations of a pretrained encoder are already speaker-blind, the scrubbing step adds nothing. The evidence in hand only partly answers this. Tomashenko et al. show identity concentrated in early layers, but a transition toward phonetic features does not mean identity is absent at depth, and rhythm survives any per-frame representation. The unscrubbed-freezing ablation at every cut depth will settle it. If it matches the scrubbed encoder, we report the depth-versus-EER-versus-WER curve and the certification procedure as the contribution, and drop the scrubber.
Why not just add differential privacy? DP bounds a different quantity, the influence of one user on the model, and the files show it is costly outside huge populations. Pro-Ind's study reports accuracy falling below 55% under strong budgets, and the ASR recipe needed about 70 million users to keep the WER cost small. Our experiment compares WER at matched attack EER in small cohorts. We also test DP layered on top of our encoder, which may be the best combination.
Why not just anonymize the audio on the device? It is a real competitor for privacy. Its weaknesses are cost and residual leakage: neural re-synthesis on every phone, residual EER of 26–37% in VoicePrivacy, and demonstrated inversion by informed attackers. We run it head-to-head against the same gradient attacks and report device compute alongside EER and WER.
Why not just run many local steps? The survey in A Survey on Gradient Inversion reports that ten or more local steps cut inversion success by over 60%. The indicator-dataset attack, however, needs no gradient matching at all and still reaches 0.86% EER. We include multi-step training as a baseline against both attack types.
Decisive Checks
We run federated recognition on Common Voice, with one client per speaker and cohort sizes of hundreds to thousands, and on TED-LIUM 3, with LibriSpeech as the public seed. The method succeeds if three things hold together. First, every gradient-level and model-inspection attack yields an EER no better than the offline probe's. Second, that EER is above 45%. Third, the WER cost over undefended training is below the small-cohort DP recipe at equal privacy.
Two outcomes would falsify the proposal. One is an attack that beats the offline probe, which would mean the bound is not operationally tight. The other is a WER cost from scrubbing that exceeds what DP pays at the same speaker EER.
References
- On the Encoding of Gender in Transformer-based ASR Representations
- How to Recover Long Audio Sequences Through Gradient Inversion Attack With Dynamic Segment-based Reconstruction
- Privacy Attacks for Automatic Speech Recognition Acoustic Models in A Federated Learning Framework
- EASY: Emotion-aware Speaker Anonymization via Factorized Distillation
- Analysis of Speech Temporal Dynamics in the Context of Speaker Verification and Voice Anonymization
- Evaluating Voice Conversion-Based Privacy Protection against Informed Attackers
- Concealing Sensitive Samples against Gradient Leakage in Federated Learning
- Mixed Quantization Enabled Federated Learning to Tackle Gradient Inversion Attacks
- General or Specific? Investigating Effective Privacy Protection in Federated Learning for Speech Emotion Recognition
- Federated Learning with Differential Privacy for End-to-End Speech Recognition
- Introducing the VoicePrivacy Initiative
- A Survey on Gradient Inversion: Attacks, Defenses and Future Directions