Computer Scientist @SRI_Intl security, privacy, ML, federated learning, GANs, chess, Formula 1
New York, USA
Joined April 2012
- Tweets534
- Following728
- Followers205
- Likes1.7K
Pinned Tweet
Recently, I had the pleasure of discussing about #PassGPT with some of my colleagues at @SRI_Intl In the article attached, you may find more details on the story: sri.com/press/story/sri-rese…
@javirandor @fperezcruz @SDSCdatascience
Looking forward to @ACSAC_Conf Don’t miss our talk on DNN watermarking! #ACSAC2024
The first #ACSAC2024 #PaperPreview today is TATTOOED by Pagnotta et al., a #DNN #watermarking technique that allows for reliable DNN ownership verification, even in extreme adversarial settings: openconf.org/acsac2024/modul…
#AI #cybersecurity
@SRI_Intl @BrilandHitaj @fperezcruz
ALT Pagnotta et al.'s "TATTOOED: A Robust Deep Neural Network Watermarking Scheme based on Spread-Spectrum Channel Coding"
Briland Hitaj retweeted
Another entry in a long-running series where Nicholas Carlini breaks ML defenses published at top security conferences with as little effort as possible (in this case a one line bugfix in the eval)
I am still in shock about this terrible news! I’ve worked with Sophia on so many projects and I was looking forward to many more to come! Sophia was one of the brightest people I have ever met! She will be missed!
Rest in peace!
Briland Hitaj retweeted
New (2h13m 😅) lecture: "Let's build the GPT Tokenizer"
Tokenizers are a completely separate stage of the LLM pipeline: they have their own training set, training algorithm (Byte Pair Encoding), and after training implement two functions: encode() from strings to tokens, and decode() back from tokens to strings. In this lecture we build from scratch the Tokenizer used in the GPT series from OpenAI.
Briland Hitaj retweeted
🧵 Can data poisoning and RLHF be combined to unlock a universal jailbreak backdoor in LLMs?
Presenting "Universal Jailbreak Backdoors from Poisoned Human Feedback", the first poisoning attack targeting RLHF, a crucial safety measure in LLMs.
📖 Paper: arxiv.org/abs/2311.14455
Earlier today, @javirandor did a fantastic presentation at #ESORICS2023 of our recent #PassGPT paper!
If you are attenting ESORICS, make sure you meet and have a chat with Javier on #LLMs, #security, and #privacy
Link to the paper: arxiv.org/pdf/2306.01545.pdf
Briland Hitaj retweeted
When analyzing ML security and privacy you need to study 𝐬𝐲𝐬𝐭𝐞𝐦𝐬, not just models!
Our new paper shows that privacy is way worse when models are deployed in systems that use data cleaners, output filters, etc.
Paper: arxiv.org/abs/2309.05610
Blog: spylab.ai/blog/side-channels…
Briland Hitaj retweeted
Love is the solution to AI alignment.
We had a nice chat with Nate on #MaliciousLife podcast discussing some of the risks with current state of the art models, especially discussing threats that may come from using unvetted pre-trained models from third party repositories!
podcasts.apple.com/us/podcas…
The #MaliciousLife podcast shorturl.at/fsD67 has highlighted our MaleficNet shorturl.at/pvCDT on their latest episode on Is Generative AI Dangerous?
We thank Nate Nelson for reaching out to us and having a challenging conversation about this paper. @BrilandHitaj
Briland Hitaj retweeted
PassGPT has been accepted for #ESORICS2023! Looking forward to exciting discussions on how LLMs can enhance security in password-based systems.
🔐 Introducing PassGPT🔓
Trained on password leaks, PassGPT can generate 20% more unseen passwords than existing GAN methods.
📖 arxiv.org/abs/2306.01545
Joint work with @fperezcruz and @BrilandHitaj.
🧵 Let's dive into our key contributions. 🧵
Briland Hitaj retweeted
😂🤣 Where do find these @docmilanfar Absolute gold mine.
🔥FedComm accepted at TDSC 🔥
Our paper "FedComm: Federated Learning as a Medium for Covert Communication" was accepted at the Transactions on Dependable and Secure Computing journal.
You can read more about our results here: arxiv.org/abs/2201.08786
In the paper we study and demonstrate the possibility of building covert communication channels on top of federated learning without hindering the performance of the model.
This is a joint work with @dorjanhitaj93, Giulio Pagnotta, @fperezcruz and Luigi V. Mancini
@SRI_Intl @SDSCdatascience @SapienzaRoma