Current LLMs are generally aligned to follow safety requirements and tend to refuse toxic prompts. However, LLMs can fail to refuse toxic prompts or be overcautious and refuse benign examples. In addition, state-of-the-art toxicity detectors have low TPRs at low FPR, incurring high costs in real-world applications where toxic examples are rare. In this paper, we introduce Moderation Using LLM Introspection (MULI), which detects toxic prompts using the information extracted directly from LLMs themselves. We found we can distinguish between benign and toxic prompts from the distribution of the first response token's logits. Using this idea, we build a robust detector of toxic prompts using a sparse logistic regression model on the first response token logits. Our scheme outperforms SOTA detectors under multiple metrics.

Ethereum is a popular blockchain that facilitates the creation of decentralized applications (dApps) and enables digital transactions to be executed without the need for a central authority. However, as in traditional markets, information asymmetry and market inefficiencies are used to the detriment of ordinary users via trading strategies that exploit “Miner Extractable Value” (MEV). We propose two extensions of Ethereum, one for proof of work (PoW), and one for proof of stake (PoS), that eliminate most forms of MEV by randomizing the execution order of transactions and hiding the content of transactions until their inclusion in a block. We simulate attack scenarios for both settings and provide detailed security properties and proofs.

Real-world adversarial attacks on machine learning models often feature an asymmetric structure wherein adversaries only attempt to induce false negatives (e.g., classify a spam email as not spam). We formalize the asymmetric robustness certification problem and correspondingly present the feature-convex neural network architecture, which composes an input-convex neural network (ICNN) with a Lipschitz continuous feature map in order to achieve asymmetric adversarial robustness. We consider the aforementioned binary setting with one "sensitive" class, and for this class we prove deterministic, closed-form, and easily-computable certified robust radii for arbitrary $\ell_p$-norms. We theoretically justify the use of these models by characterizing their decision region geometry, extending the universal approximation theorem for ICNN regression to the classification setting, and proving a lower bound on the probability that such models perfectly fit even unstructured uniformly distributed data in sufficiently high dimensions. Experiments on Malimg malware classification and subsets of the MNIST, Fashion-MNIST, and CIFAR-10 datasets show that feature-convex classifiers attain substantial certified $\ell_1$, $\ell_2$, and $\ell_{\infty}$-radii while being far more computationally efficient than competitive baselines.

When employing supervised machine learning to analyze network traffic, the heart of the task often lies in developing effective features for the ML to leverage. We develop GGFAST, a unified, automated framework that can build powerful classifiers for specific network traffic analysis tasks, built on interpretable features. The framework uses only packet sizes, directionality, and sequencing, facilitating analysis in a payload-agnostic fashion that remains applicable in the presence of encryption.