Hacking Back the AI-Hacker: Prompt Injection as a Defense Against LLM-driven Cyberattacks

Large language models (LLMs) are increasingly being harnessed to automate cyberattacks, making sophisticated exploits more accessible and scalable. In response, we propose a new defense strategy tailored to counter LLM-driven cyberattacks. We introduce Mantis, a defensive framework that exploits LLMs' susceptibility to adversarial inputs to undermine malicious operations. Upon detecting an automated cyberattack, Mantis plants carefully crafted inputs into system responses, leading the attacker's LLM to disrupt their own operations (passive defense) or even compromise the attacker's machine (active defense). By deploying purposefully vulnerable decoy services to attract the attacker and using dynamic prompt injections for the attacker's LLM, Mantis can autonomously hack back the attacker. In our experiments, Mantis consistently achieved over 95% effectiveness against automated LLM-driven attacks. To foster further research and collaboration, Mantis is available as an open-source tool: https://github.com/pasquini-dario/project_mantis

LLMmap: Fingerprinting For Large Language Models

We introduce LLMmap, a first-generation fingerprinting attack targeted at LLM-integrated applications. LLMmap employs an active fingerprinting approach, sending carefully crafted queries to the application and analyzing the responses to identify the specific LLM model in use. With as few as 8 interactions, LLMmap can accurately identify LLMs with over 95% accuracy. More importantly, LLMmap is designed to be robust across different application layers, allowing it to identify LLMs operating under various system prompts, stochastic sampling hyperparameters, and even complex generation frameworks such as RAG or Chain-of-Thought.

Watermarks in the Sand: Impossibility of Strong Watermarking for Generative Models

Watermarking generative models consists of planting a statistical signal (watermark) in a model's output so that it can be later verified that the output was generated by the given model. A strong watermarking scheme satisfies the property that a computationally bounded attacker cannot erase the watermark without causing significant quality degradation. In this paper, we study the (im)possibility of strong watermarking schemes. We prove that, under well-specified and natural assumptions, strong watermarking is impossible to achieve. This holds even in the private detection algorithm setting, where the watermark insertion and detection algorithms share a secret key, unknown to the attacker. To prove this result, we introduce a generic efficient watermark attack; the attacker is not required to know the private key of the scheme or even which scheme is used. Our attack is based on two assumptions: (1) The attacker has access to a "quality oracle" that can evaluate whether a candidate output is a high-quality response to a prompt, and (2) The attacker has access to a "perturbation oracle" which can modify an output with a nontrivial probability of maintaining quality, and which induces an efficiently mixing random walk on high-quality outputs. We argue that both assumptions can be satisfied in practice by an attacker with weaker computational capabilities than the watermarked model itself, to which the attacker has only black-box access. Furthermore, our assumptions will likely only be easier to satisfy over time as models grow in capabilities and modalities. We demonstrate the feasibility of our attack by instantiating it to attack three existing watermarking schemes for large language models: Kirchenbauer et al. (2023), Kuditipudi et al. (2023), and Zhao et al. (2023). The same attack successfully removes the watermarks planted by all three schemes, with only minor quality degradation.

Universal Neural-Cracking-Machines: Self-Configurable Password Models from Auxiliary Data

We develop the first universal password model -- a password model that, once pre-trained, can automatically adapt to any password distribution. To achieve this result, the model does not need to access any plaintext passwords from the target set. Instead, it exploits users' auxiliary information, such as email addresses, as a proxy signal to predict the underlying target password distribution. The model uses deep learning to capture the correlation between the auxiliary data of a group of users (e.g., users of a web application) and their passwords. It then exploits those patterns to create a tailored password model for the target community at inference time. No further training steps, targeted data collection, or prior knowledge of the community's password distribution is required. Besides defining a new state-of-the-art for password strength estimation, our model enables any end-user (e.g., system administrators) to autonomously generate tailored password models for their systems without the often unworkable requirement of collecting suitable training data and fitting the underlying password model. Ultimately, our framework enables the democratization of well-calibrated password models to the community, addressing a major challenge in the deployment of password security solutions on a large scale.

Eluding Secure Aggregation in Federated Learning via Model Inconsistency

Federated learning allows a set of users to train a deep neural network over their private training datasets. During the protocol, datasets never leave the devices of the respective users. This is achieved by requiring each user to send "only" model updates to a central server that, in turn, aggregates them to update the parameters of the deep neural network. However, it has been shown that each model update carries sensitive information about the user's dataset (e.g., gradient inversion attacks). The state-of-the-art implementations of federated learning protect these model updates by leveraging secure aggregation: A cryptographic protocol that securely computes the aggregation of the model updates of the users. Secure aggregation is pivotal to protect users' privacy since it hinders the server from learning the value and the source of the individual model updates provided by the users, preventing inference and data attribution attacks. In this work, we show that a malicious server can easily elude secure aggregation as if the latter were not in place. We devise two different attacks capable of inferring information on individual private training datasets, independently of the number of users participating in the secure aggregation. This makes them concrete threats in large-scale, real-world federated learning applications. The attacks are generic and do not target any specific secure aggregation protocol. They are equally effective even if the secure aggregation protocol is replaced by its ideal functionality that provides the perfect level of security. Our work demonstrates that secure aggregation has been incorrectly combined with federated learning and that current implementations offer only a "false sense of security".

Unleashing the Tiger: Inference Attacks on Split Learning

We investigate the security of split learning -- a novel collaborative machine learning framework that enables peak performance by requiring minimal resources consumption. In the paper, we make explicit the vulnerabilities of the protocol and demonstrate its inherent insecurity by introducing general attack strategies targeting the reconstruction of clients' private training sets. More prominently, we demonstrate that a malicious server can actively hijack the learning process of the distributed model and bring it into an insecure state that enables inference attacks on clients' data. We implement different adaptations of the attack and test them on various datasets as well as within realistic threat scenarios. To make our results reproducible, we made our code available at https://github.com/pasquini-dario/SplitNN_FSHA.

Reducing Bias in Modeling Real-world Password Strength via Deep Learning and Dynamic Dictionaries

Password security hinges on an accurate understanding of the techniques adopted by attackers. However, current studies mostly rely on probabilistic password models that are imperfect proxies of real-world guessing strategies. The main reason is that attackers rely on very pragmatic approaches such as dictionary attacks. Unfortunately, it is inherently difficult to correctly model those methods. To be representative, dictionary attacks must be thoughtfully configured according to a process that requires an expertise that cannot be easily replicated in password studies. The consequence of inaccurately calibrating those attacks is the unreliability of password security estimates, impaired by measurement bias. In the present work, we introduce new guessing techniques that make dictionary attacks consistently more resilient to inadequate configurations. Our framework allows dictionary attacks to self-heal and converge towards optimal attacks' performance, requiring no supervision or domain-knowledge. To achieve this: (1) We use a deep neural network to model and then simulate the proficiency of expert adversaries. (2) Then, we introduce automatic dynamic strategies within dictionary attacks to mimic experts' ability to adapt their guessing strategies on the fly by incorporating knowledge on their targets. Our techniques enable robust and sound password strength estimates, eventually reducing bias in modeling real-world threats in password security.

Interpretable Probabilistic Password Strength Meters via Deep Learning

Probabilistic password strength meters have been proved to be the most accurate tools to measure password strength. Unfortunately, by construction, they are limited to solely produce an opaque security estimation that fails to fully support the user during the password composition. In the present work, we move the first steps towards cracking the intelligibility barrier of this compelling class of meters. We show that probabilistic password meters inherently own the capability of describing the latent relation occurring between password strength and password structure. In our approach, the security contribution of each character composing a password is disentangled and used to provide explicit fine-grained feedback for the user. Furthermore, unlike existing heuristic constructions, our method is free from any human bias, and, more importantly, its feedback has a clear probabilistic interpretation. In our contribution: (1) we formulate the theoretical foundations of interpretable probabilistic password strength meters; (2) we describe how they can be implemented via an efficient and lightweight deep learning framework suitable for client-side operability.

Deep Models Under the GAN: Information Leakage from Collaborative Deep Learning

Deep Learning has recently become hugely popular in machine learning, providing significant improvements in classification accuracy in the presence of highly-structured and large databases. Researchers have also considered privacy implications of deep learning. Models are typically trained in a centralized manner with all the data being processed by the same training algorithm. If the data is a collection of users' private data, including habits, personal pictures, geographical positions, interests, and more, the centralized server will have access to sensitive information that could potentially be mishandled. To tackle this problem, collaborative deep learning models have recently been proposed where parties locally train their deep learning structures and only share a subset of the parameters in the attempt to keep their respective training sets private. Parameters can also be obfuscated via differential privacy (DP) to make information extraction even more challenging, as proposed by Shokri and Shmatikov at CCS'15. Unfortunately, we show that any privacy-preserving collaborative deep learning is susceptible to a powerful attack that we devise in this paper. In particular, we show that a distributed, federated, or decentralized deep learning approach is fundamentally broken and does not protect the training sets of honest participants. The attack we developed exploits the real-time nature of the learning process that allows the adversary to train a Generative Adversarial Network (GAN) that generates prototypical samples of the targeted training set that was meant to be private (the samples generated by the GAN are intended to come from the same distribution as the training data). Interestingly, we show that record-level DP applied to the shared parameters of the model, as suggested in previous work, is ineffective (i.e., record-level DP is not designed to address our attack).

Hacking Smart Machines with Smarter Ones: How to Extract Meaningful Data from Machine Learning Classifiers

Machine Learning (ML) algorithms are used to train computers to perform a variety of complex tasks and improve with experience. Computers learn how to recognize patterns, make unintended decisions, or react to a dynamic environment. Certain trained machines may be more effective than others because they are based on more suitable ML algorithms or because they were trained through superior training sets. Although ML algorithms are known and publicly released, training sets may not be reasonably ascertainable and, indeed, may be guarded as trade secrets. While much research has been performed about the privacy of the elements of training sets, in this paper we focus our attention on ML classifiers and on the statistical information that can be unconsciously or maliciously revealed from them. We show that it is possible to infer unexpected but useful information from ML classifiers. In particular, we build a novel meta-classifier and train it to hack other classifiers, obtaining meaningful information about their training sets. This kind of information leakage can be exploited, for example, by a vendor to build more effective classifiers or to simply acquire trade secrets from a competitor's apparatus, potentially violating its intellectual property rights.