cd /news/ai-safety/eurecom-finds-adding-voice-spoofing-… · home › topics › ai-safety › article
[ARTICLE · art-171457] src=alphasignal.ai ↗ pub= topic=ai-safety verified=true sentiment=· neutral

EURECOM Finds Adding Voice Spoofing Data Makes AI Detectors 20% Worse

Researchers at EURECOM and the Laboratoire d'informatique d'Avignon found that pooling multiple speech-spoofing corpora can degrade deepfake detectors, and their Invariant Domain Feature Extraction (IDFE) framework cuts pooled equal error rate by up to 20%, from 6.11% to 4.88%, against the MHFA baseline when training on ASVspoof 5 plus 2019. IDFE adds a corpus classifier behind a gradient reversal layer so the encoder learns to ignore which dataset a clip came from; adding Fake-or-Real as a third corpus drops pooled EER from 5.48% to 4.69%, and ASVspoof 2021 LA improves 31% in Case 3. The paper reports limits: ASVspoof 5 still performs best when trained alone, results are sensitive to the α parameter, and no code release is mentioned.

recommendation

When pooling spoofing corpora, add a domain-adversarial corpus classifier with gradient reversal to the shared encoder, and tune α, since results are sensitive to it.

more in playbook: Fine-tuning and data
read3 min views5 publishedNov 18, 2026
EURECOM Finds Adding Voice Spoofing Data Makes AI Detectors 20% Worse
Image: Alphasignal (auto-discovered)

Pooling more training corpora can make speech deepfake detectors worse. IDFE adds a corpus classifier with gradient reversal so the model can't lean on dataset quirks, and it cuts pooled error by up to 20%.

  • New paper : training speech spoofing detectors on more corpora can lower accuracy because of dataset-specific biases.

  • IDFE adds a corpus classifier behind a gradient reversal layer, so the encoder learns to ignore which dataset a clip came from.

  • Training on ASVspoof 5 plus 2019, pooled EER falls 20%, from 6.11% to 4.88%, against the MHFA baseline.

  • With FoR added as a third corpus, pooled EER drops from 5.48% to 4.69%. ASVspoof 2021 LA improves 31% in Case 3.

  • Limits: ASVspoof 5 still does best when trained on alone, results are sensitive to α, and no code release is mentioned.

  • The same team's speaker-invariant follow-up applies the idea to speaker identity and reports a 25.7% EER reduction. A common assumption in speech machine learning is that adding training corpora produces a more robust model. Researchers from the Laboratoire d'informatique d'Avignon and EURECOM tested that assumption on spoofing detection, which distinguishes human speech from synthetic or converted voices. Their experiments found that multi-corpus training can raise error rates by teaching a detector to recognize datasets instead of durable spoofing cues. The team’s Invariant Domain Feature Extraction framework addresses that shortcut with a domain classifier and gradient reversal.

More corpora, worse detector #

Each spoofing corpus bundles its own microphones, codecs, recording conditions, attack generators and preprocessing choices. A detector can use those properties as shortcuts when they correlate with genuine or spoofed labels. Those correlations often fail on audio from another source, producing unstable cross-domain performance.

Equal error rate, or EER, measures the point at which the proportion of spoofed clips accepted equals the proportion of genuine clips rejected. Lower values indicate better detection. The baseline Multi-Head Factorized Attentive Pooling model, trained only on ASVspoof 5 in Case 1, records 4.64% EER on the ASVspoof 5 evaluation set and 12.14% on ASVspoof 2021 LA. Training only on ASVspoof 2019 in Case 2 shifts the weakness: EER reaches 9.33% on ASVspoof 2021 LA and 12.68% on ASVspoof 5.

ASVspoof 5 exposes the clearest negative transfer, where data from another source harms performance on the target corpus. Adding ASVspoof 2019 and Fake-or-Real raises the ASVspoof 5 EER from the 4.64% single-corpus result to 8.36%. The mixed model has seen more audio, yet corpus-specific bias makes it less accurate on that evaluation set.

How the model forgets the corpus #

Invariant Domain Feature Extraction, or IDFE, applies domain-adversarial training to the detector’s shared embedding. A self-supervised learning encoder first converts raw audio into representations learned during pretraining. The MHFA back end then uses attention to combine information from multiple encoder layers into one embedding for classification.

This story is for Pro members #

You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.

── more in #ai-safety 4 stories · sorted by recency
── more on @eurecom 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/eurecom-finds-adding…] indexed:0 read:3min 2026-11-18 · —