Training agents at @datadoghq and failing to train a cat at home. Prev. @AdaptiveML, @VaticLabs, @AmazonScience. PhD @LTIatCMU, Ingénieur @Polytechnique
New York, USA
Joined October 2021
- Tweets35
- Following198
- Followers101
- Likes75
Raphael Olivier retweeted
Excited to introduce Scout 1.0, the model powering Bits Investigation, Datadog’s AI SRE agent for complex troubleshooting and late-night alerts.
We post-trained GLM-5.3 Flash to match frontier models at much lower cost on our internal AlertEval benchmark. 🧵
Happy to announce that our paper "How many perturbations break this model? Evaluating robustness beyond adversarial accuracy" will be presented at ICML 2023!
1/9 In adversarial robustness, we measure performance with accuracy. We attack an input; if the attack is successful we assign 0, if not 1; then we average over the test set.
This approach has flaws, which we start addressing in our new preprint: arxiv.org/abs/2207.04129
Whisper is famously robust, but is it adversarially robust?
The answer in our new preprint: arxiv.org/abs/2210.17316
Code github.com/RaphaelOlivier/wh…
Adversarial examples are on🤗Datasets huggingface.co/datasets/Raph…
1/6
There are immediate applications of that kind of attacks. If you use ASR models to transcribe online multimedia data, any malicious party can upload adversarial examples, ready for the taking. They could for instance evade censorship - or trigger it.
5/6
ASR may achieve impressive results these days, but there is a long way to go before it can be safely used in all contexts.
We released our adversarial examples as a🤗dataset (huggingface.co/datasets/Raph…). You can listen to them online, or try them out on Whisper models locally.
6/6
1/9 In adversarial robustness, we measure performance with accuracy. We attack an input; if the attack is successful we assign 0, if not 1; then we average over the test set.
This approach has flaws, which we start addressing in our new preprint: arxiv.org/abs/2207.04129
8/9 The latter is already known for [augmentation+adversarial training]... but only since last year (arxiv.org/abs/2103.01946). Before it was thought that data augmentation has no effect on robustness.
This shows how sparsity can help nudge research in the right direction.
Raphael Olivier retweeted
📢 SPSC Webinar: Targeted and transferable adversarial perturbations on self-supervised ASR models
🗓️ Mon. August 1st, 16:00 CEST, by @rapholivie (Carnegie Mellon University)
✒️ Register here:
spsc-sig.org/form/webinar-re…
It's free & open to everyone!
Our paper "Recent improvements of ASR models in the face of adversarial attacks" was accepted @ISCAInterspeech 2022!
One of our contributions in this work was the robust_speech package, so I'll take this opportunity to present some features we recently added in version 0.2
1/9 Happy to release our robustness package for Speech Recognition models, robust_speech, and its accompanying preprint:
github.com/RaphaelOlivier/ro…
arxiv.org/abs/2203.16536
-New attacks: Yeehaw Junction (arxiv.org/abs/2203.05408), universal audio attack (arxiv.org/abs/1905.03828)
-We ported our very own Sequential randomized smoothing defense (arxiv.org/abs/2112.03000)
-CommonVoice and SpeechCommand datasets are now natively supported
-@fairseq Wav2Vec2 models are supported (on top of @huggingface's)
-And various bug fixes.
If you work on speech security, give it a try!