A backdoor attack plants a hidden behavior in a model during training: the model looks and works normally until it hears a specific trigger, at which point it does whatever the attacker wants. The security community has built run-time defenses on a quiet but load-bearing assumption — that a backdoor trigger is stuck inside the poisoned model's weights, and so will look wrong to any other, independent model asked to judge the same input. Meanwhile, a separate literature on adversarial transferability has shown almost the opposite: small input perturbations can be deliberately optimized to fool many models at once, including ones the attacker never saw. This proposal asks what happens when those two facts collide: can an attacker craft an audio backdoor trigger that, in addition to activating the poisoned model, deliberately transfers into the representations of independent clean feature extractors — and, while doing so, stays below the psychoacoustic masking thresholds of human hearing and survives the perturbation-based detectors designed for exactly this threat?
The stake is concrete. The non-transferability defense NTD: Non-Transferability Enabled Backdoor Detection flags an input as backdoored precisely by measuring a representational gap between the poisoned model's prediction and a clean feature extractor's view of the input. If a trigger can be optimized so that clean feature extractors see triggered audio as an ordinary member of the attacker's target class, that defense's core signal disappears — and the same optimization machinery gives the trigger the sample-specific, imperceptible character that has already let compression-based backdoors slip past STRIP-style detectors. We propose to build and study such triggers in the speech domain, and to test the strongest existing audio backdoor defenses against them in a properly adaptive evaluation.
Core Claim
The claim is that backdoor trigger non-transferability is a property of how triggers have been built so far, not of backdoors as a mechanism. Prior audio triggers are either static signal-processing effects — Going in Style: Audio Backdoors Through Stylistic Transformations applies guitar-pedal-style transformations like pitch shifting and reverb as triggers, achieving up to 96% attack success on keyword spotting with only a 0.24% average drop in clean accuracy — or sample-specific compression fusions optimized against a single surrogate, as in CBA: Backdoor Attack on Deep Speech Classification via Audio Compression. Neither ever asks whether the trigger's representation lands where a clean feature extractor would expect the target class to land. We hypothesize that adding exactly that term — an ensemble of independent feature extractors pulled into the trigger's optimization objective, in the spirit of the Dynamic Normalized Ensemble that lets DUAP: Dual-task Universal Adversarial Perturbations Against Voice Control Systems transfer single perturbations across architectures — produces triggers that (i) activate a poisoned victim at CBA-level success rates, (ii) are invisible to a non-transferability detector ported to audio, (iii) stay below psychoacoustic masking thresholds rather than merely below a loss-based perceptual penalty, and (iv) survive, partially or fully, the entropy-based detection and time-frequency purification of SpeechGuard: Online Defense against Backdoor Attacks on Speech Recognition Models.
This is a study of a defensive axiom under adaptive pressure, not just a new attack. The methodological lens comes from Optimizing Adaptive Attacks against Content Watermarks for Language Models, which shows that static evaluations systematically understate risk against defense-aware adversaries; our experiments treat every defense as a known, white-box component of the attacker's training loop.
Research Program
We build a trigger generator for keyword-spotting and speaker-identification models, following CBA's sample-specific mask-fusion template but with two new ingredients. The generator learns, per input $x$, a soft mask $v$ that fuses transformed and original acoustic features to produce a triggered input $\hat{x}$. The optimization is
$$ \min_{v}\;\; \mathcal{L}_{\text{atk}}(\hat{x}) \;+\; \lambda \sum_{f \in \mathcal{F}} d\!\left(f(\hat{x}),\, \mu_{f, y_t}\right) \quad \text{subject to} \quad |\hat{x} - x| \preceq \tau(x), $$
where $\mathcal{L}_{\text{atk}}$ is the usual poisoning loss that makes the victim classify $\hat{x}$ as the attacker's target class $y_t$, $\mathcal{F}$ is an ensemble of independent, clean, publicly available feature extractors (frozen — none of them is the victim), $\mu_{f, y_t}$ is the centroid of class $y_t$ in extractor $f$'s representation space, $d$ is a distance in that space, and $\tau(x)$ is the per-time-frequency-bin psychoacoustic masking threshold of the input, so every perturbation bin sits below the level a human ear can detect given the masking energy of the surrounding speech. The second term is the proposal's key ingredient: it does for backdoor triggers what ensemble-based transfer attacks do for evasion perturbations — it optimizes for cross-model representational agreement instead of overfitting one model. Ensembling across heterogeneous extractors follows An Adaptive Model Ensemble Adversarial Attack for Boosting Adversarial Transferability, which shows that balancing losses across surrogates stabilizes gradients across unseen architectures. We additionally include an entropy-maintenance term that maximizes the victim's prediction entropy on mixtures of $\hat{x}$ with SNR-calibrated noise, making the trigger adaptive to S-STRIP's exact detection mechanism rather than merely hard for it in practice.
Three analyses accompany the build. First, a representational study of why the trick works or fails: following the cross-representation findings of D-CAPTCHA++: A Study of Resilience of Deepfake CAPTCHA under Transferable Imperceptible Adversarial Attack. — where attacks crafted on LFCC-feature models transferred well to other LFCC models but poorly to raw-waveform models like RawNet2 — we measure whether ensemble optimization can push a trigger across the spectral-to-raw-waveform boundary that currently blocks evasion transferability. Second, a defense-signal analysis: we trace NTD's discrepancy score and S-STRIP's entropy statistic for our triggers versus CBA and JingleBack triggers, to show which defensive signal each ingredient removes. Third, a perceptual audit with psychoacoustic threshold compliance and listener studies, since loss-based perceptual penalties have documented failure modes: Audio Adversarial Examples: Targeted Attacks on Speech-to-Text shows that low-energy noise in quiet segments is heard as hiss, and Adversarial Attacks Against Automatic Speech Recognition Systems via Psychoacoustic Hiding established masking thresholds as the principled fix — a fix that, per the corpus, has never been applied to training-time backdoor triggers, only inference-time evasion.
The threat model is poisoning with zero query access to any deployed system: the attacker controls a fraction of the training data and nothing else. This is the regime where transfer attacks are the only option, since query-based optimization of the kind in Query-Based Adversarial Prompt Generation requires live feedback that a poisoning-time adversary does not have.
Proposed Experiments
We evaluate on Google Speech Commands v2 (the 10-class, 21,115-sample setup used by CBA, JingleBack, and SpeechGuard) and AudioMNIST, with TIMIT for a speaker-identification extension following JingleBack. Victim models span feature front-ends deliberately: ResNet34 on spectrograms (CBA's strongest reported setting), an LFCC-based model, and a raw-waveform network, so cross-representation transfer is measured rather than assumed. Independent feature extractors $\mathcal{F}$ for both the attack ensemble and the audio port of NTD come from different pretraining lineages than the victims, mirroring NTD's clean-feature-extractor design. Poisoning rates follow CBA's protocol (around 5%). Metrics are attack success rate (ASR) on triggered inputs, clean accuracy (ACC), the audio port of NTD's discrepancy score, S-STRIP entropy distributions and false-acceptance/false-rejection rates, post-purification ASR under SpeechGuard's autoencoder time-frequency masking, and psychoacoustic threshold compliance plus a small mean-opinion-score listening study. The decisive comparisons are: our trigger versus CBA under a clean-data-independent NTD-style detector; our trigger versus static triggers under SpeechGuard at matched clean-accuracy budgets; and ablations dropping the feature-extractor ensemble term and the masking constraint separately, which isolate the two ingredients' contributions to detector evasion and imperceptibility respectively.
Baselines
The do-nothing reference point is an undefended poisoned model: a standard trigger such as JingleBack's stylistic transformations is inserted with no run-time defense, which on Google Speech Commands and TIMIT achieves attack success up to 96% with an average clean-accuracy drop of only 0.24% (Going in Style: Audio Backdoors Through Stylistic Transformations); any serious attack must at least match this stealth-accuracy trade-off before defenses even enter the picture. The strongest existing attack competitor is CBA, the sample-specific compression-fusion backdoor, which on Google Speech Commands at a 5% poisoning rate exceeds 99% attack success — 99.73% ASR with 96.69% clean accuracy on ResNet34 — and already evades STRIP, Neural Cleanse, filtering, and pruning (CBA: Backdoor Attack on Deep Speech Classification via Audio Compression). We run CBA with its published optimization settings on identical data splits, victims, and poisoning budgets, giving it the same compute and tuning effort as our method. Beating CBA means matching its ASR and clean accuracy while additionally passing the audio-ported NTD detector that flags CBA — CBA's trigger was never optimized to transfer into clean feature extractors, so its discrepancy score should be high where ours is low. The strongest defense baseline is SpeechGuard, whose S-STRIP stage achieves a false-acceptance rate below 10% in most scenarios and whose purification stage cuts attack success by over 90% across static trigger types and models, at the cost of clean-accuracy declines of up to 10% (SpeechGuard: Online Defense against Backdoor Attacks on Speech Recognition Models); we run it as published, on the same victims and data, tuning only its published SNR and mask parameters within its stated ranges. Beating SpeechGuard means keeping attack success materially above the post-defense level it imposes on static triggers — we pre-register "above 50% ASR after full detection-plus-purification" as success — while its clean-accuracy cost stays as reported. Finally, the key-ingredient ablation is our own trigger with the feature-extractor ensemble term removed (single-surrogate optimization, everything else unchanged): if the ablated version is caught by the audio NTD port as reliably as CBA is, the transfer term is doing the work; if the ablated version also passes, our claim about why the defense breaks is wrong and we will report that.
Why not just…?
Why not just bound the perturbation with an $L_p$ norm or reuse standard compression, skipping the psychoacoustic modeling? The corpus gives direct evidence this fails perceptually: in DUAP: Dual-task Universal Adversarial Perturbations Against Voice Control Systems, a universal perturbation crafted without psychoacoustic masking degraded audio to an SNR of -21.8 dB — unintelligible to listeners — while the masked version held -6.96 dB with good opinion scores; and Going in Style reports that several of its unconstrained stylistic triggers remained clearly audible. CBA's compression triggers constrain imperceptibility only through a general perceptual loss and spectrogram inspection, with no biological masking bound. Our ablation without the $\tau(x)$ constraint will quantify the gap directly on the same trigger generator, so the claim is testable rather than rhetorical.
Why not just assume fine-pruning or other static defenses will remove the backdoor, making adaptive trigger design unnecessary? This is the alternative with real teeth, and we will not argue around it: fine-pruning with clean data is a genuine threat to our attack, and our proposal does not target it. But the corpus shows static defenses are not a settled win even against non-adaptive audio backdoors — CBA reports that no pruning point reliably neutralized its backdoor, and Fine-Pruning: Defending Against Backending Attacks on Deep Neural Networks itself shows a pruning-aware attacker limiting the defense's effect (traffic-sign attack success fell only from roughly 99% to about 30%), while Adaptive Backdoor Attack against Deep Neural Networks requires over 90% of neurons to be pruned before its backdoor breaks, destroying clean accuracy. We will run fine-pruning and fine-tuning against our triggers and report the results honestly; our claim is specifically about run-time, clean-data-light defenses, and if fine-pruning also kills our trigger at acceptable clean cost, that bounds the attack's practical reach and we will say so.
Why not just use query-based adaptive attacks instead of engineering zero-query transferability? If the attacker has live query access to the target, query-based optimization is indeed more reliable — but our threat model is a poisoning-time adversary who by construction has no deployed system to query, and Query-Based Adversarial Prompt Generation documents that heavy querying is expensive and readily flagged by anomaly monitoring. The zero-query rationale is laid out in QFA2SR: Query-Free Adversarial Transfer Attacks to Speaker Recognition Systems and Zero-Query Transfer Attacks on Context-Aware Object Detectors: single-shot attacks avoid the monitor entirely. So the alternative applies to a different adversary, not a simpler solution to ours.
Why not just deploy STRIP or NTD as-is and call backdoor detection solved? STRIP's own authors flag non-static triggers as an open challenge, and CBA's sample-specific triggers already evaded STRIP and Neural Cleanse; SpeechGuard's S-STRIP inherits STRIP's input-agnostic assumptions and has only ever been tested against static triggers, while paying a clean-accuracy cost of up to 10% when it mis-fires. NTD is the more interesting case: it is built on the unverified premise that triggers cannot transfer, it has only been tested in vision — no reported result found on Lacuna for NTD on audio or on any transfer-optimized trigger — and it degrades under high intra-class variance, which speaker and keyword audio exhibits. Our experiments are precisely the test of that premise, and NTD's reported near-zero false-acceptance on FaceScrub with 0.058s detection latency (NTD: Non-Transferability Enabled Backdoor Detection) is the bar our audio port must first reproduce on static triggers before we claim anything about breaking it.
Decisive Checks
The proposal stands or falls on three outcomes. First, under an audio port of NTD validated against static triggers (it must flag JingleBack and CBA triggers at low false-acceptance before it counts as a defense at all), our transfer-optimized trigger should be statistically indistinguishable from clean audio in discrepancy score while the ablated single-surrogate version is flagged — that contrast is the central result. Second, under SpeechGuard run as published, our entropy-adapted trigger should retain materially higher attack success than static triggers do after the reported over-90% ASR reduction; if purification alone strips it, the honest conclusion is that time-frequency masking is the real bottleneck and the adaptive frontier moves there. Third, cross-representation attack success from spectrogram surrogates onto raw-waveform victims should be measured and reported either way, because D-CAPTCHA++'s finding that RawNet resists LFCC-crafted transfers is the strongest existing evidence that our ensemble trick may simply fail across feature boundaries — and if it does, that failure itself would sharpen, rather than soften, the boundary conditions of the non-transferability axiom the whole question turns on.
References
- NTD: Non-Transferability Enabled Backdoor Detection
- Going in Style: Audio Backdoors Through Stylistic Transformations
- CBA: Backdoor Attack on Deep Speech Classification via Audio Compression
- DUAP: Dual-task Universal Adversarial Perturbations Against Voice Control Systems
- SpeechGuard: Online Defense against Backdoor Attacks on Speech Recognition Models
- Optimizing Adaptive Attacks against Content Watermarks for Language Models
- An Adaptive Model Ensemble Adversarial Attack for Boosting Adversarial Transferability
- D-CAPTCHA++: A Study of Resilience of Deepfake CAPTCHA under Transferable Imperceptible Adversarial Attack.
- Audio Adversarial Examples: Targeted Attacks on Speech-to-Text
- Adversarial Attacks Against Automatic Speech Recognition Systems via Psychoacoustic Hiding
- Query-Based Adversarial Prompt Generation
- Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks
- Adaptive Backdoor Attack against Deep Neural Networks
- QFA2SR: Query-Free Adversarial Transfer Attacks to Speaker Recognition Systems
- Zero-Query Transfer Attacks on Context-Aware Object Detectors