2025-02-06
Transfer Attacks Reveal SLM Vulnerabilities and Effective Noise Defenses
This section evaluates transfer attacks on SLMs by applying adversarial perturbations from a surrogate model to victim models. FlanT5-based SLMs show more robustness in grey-box settings, but black-box experiments reveal...
Cross-Prompt Attacks and Data Ablations Impact SLM Robustness
This appendix section covers three experiments: (A.3) Cross-prompt attacks, using 10 random perturbations per target, are less effective than sample-specific ones but still beat random noise; (A.4) Training data ablations...
Safety Alignment and Jailbreak Attacks Challenge Modern LLMs
This section reviews the background on safety alignment for LLMs, highlighting the HHH criteria and red team prompt crafting used to deter harmful outputs. It examines both manual and automated...
Audio Encoder Pre-training and Evaluation Enhance SLM Safety
This appendix details the pre-training and evaluation of our audio encoder for speech language models (SLMs). The encoder is a 24-layer Conformer with 300M parameters pre-trained using the BEST-RQ method...
Integrated Speech Language Models Face Critical Safety Vulnerabilities
This study examines the safety alignment of speech language models in Spoken QA. It shows that adversaries with white-box access can jailbreak these systems using nearly imperceptible perturbations, and that...
SpeechVerse Unites Audio Encoder and LLM for Superior Spoken QA
This section details the experimental setup for SpeechVerse, our unified speech language model. It describes using a 24-layer Conformer audio encoder paired with two LLMs—Flan-T5-XL and Mistral-7B variants—for spoken QA....
Unified Speech and Language Models Can Be Vulnerable to Adversarial Attacks
This study investigates the safety and robustness of integrated speech and language models (SLMs) that follow speech instructions. It demonstrates that adversarial attacks—both white-box and transfer-based—can jailbreak SLM safety guardrails...
SLMs Outperform Competitors Yet Suffer Rapid Adversarial Jailbreaks
This section discusses the results and insights from evaluating our SpeechVerse SLMs. Our models outperform competitors like SpeechGPT, showing over 40% better safety and 20% improved helpfulness, thanks to effective...
Applying Modern Technology to Business: Shoyu Ro’s Path to Success
Shoyu Ro’s journey from an intern to a tech entrepreneur showcases his expertise in AI and SaaS. He led major projects at AnyMind and FLUX before founding LR Inc., where...
Adversarial Settings and Random Noise Reveal Speech LLM Vulnerabilities
This section details the attack and countermeasure settings for SpeechVerse. Using a step size of 0.00001 and up to 100 iterations with early-stopping on unsafe responses, adversarial attacks are run...
Datasets and Evaluation Define the Robustness of Speech Language Models
This section describes the datasets and evaluation methods for SpeechVerse. The training data includes 2.5K hours of ASR speech-text pairs and 150 hours of Spoken QA pairs generated via TTS....
Adversarial Attacks Challenge the Integrity of Speech Language Models
This section details adversarial attacks and defenses for spoken QA in speech language models (SLMs). It explains white-box attacks using gradient-based methods like PGD and transfer attacks via cross-model and...
How To Run DeepSeek R1 Locally on Your (Even Slow) Laptop Completely Free
It took me 7 minutes to set up on my Macbook.Continue reading on The Startup »
A huge list of trans Girl Scouts who are selling Girl Scout...
A huge list of trans Girl Scouts who are selling Girl Scout Cookies. “Please consider choosing a trans girl scout to get your cookies from this year — the kids...
The HackerNoon Newsletter: Space Telescopes Might Be Seeing More Than They Bargained For (2/6/2025)
How are you, hacker? 🪐 What’s happening in tech today, February 6, 2025? The HackerNoon Newsletter brings the HackerNoon homepage straight to your inbox. On this day, we present you...
HackerNoon Decoded 2024: Celebrating Our Startups Community!
Welcome to HackerNoon Decoded—the ultimate recap of the Startups' stories, writers, and trends that defined 2024! Explore the top Startups' stories that captivated our readers, meet the leading writers who...
Why Gas Fees Drop as Transaction Volumes Increase on zkSync
The analysis reveals that on L2 rollups like zkSync, gas fees tend to decrease as transaction volumes rise, due to transaction batching and compression. ZK rollups like zkSync are particularly...
Unsurprisingly, it would be bad if Elon Musk breaks into the computer...
Unsurprisingly, it would be bad if Elon Musk breaks into the computer systems of the National Nuclear Security Administration. “It has all manner of sensitive information on hand, including nuclear-weapon...
Mitigating the Effects of Photobombing in Exoplanet Observations
Photobombing from neighboring planets complicates exoplanet observations, but strategies like larger telescopes, improved detector sampling, multi-epoch observations, and advanced spectroscopy can help mitigate these issues.
How Nearby Planets Impact the Spectra of Earth-like Exoplanets
Photobombing by neighboring planets, such as the Moon or Mars, introduces noise and flux variations in the spectra of Earth-like exoplanets, making it harder to detect molecular biosignatures and leading...
Welcome to Postreads
Discover and follow the best content from across the web, all in one place. Create an account to start building your personalized feed today.
