Cryptography and Security (cs.CR)

Abstract: Widespread adoption and growing popularity of embedded/IoT/CPS devices make them attractive attack targets. On low-to-mid-range devices, security features are typically few or none due to various constraints. Such devices are thus subject to malware-based compromise. One popular defensive measure is Remote Attestation (RA) which allows a trusted entity to determine the current software integrity of an untrusted remote device. For higher-end devices, RA is achievable via secure hardware components. For low-end (bare metal) devices, minimalistic hybrid (hardware/software) RA is effective, which incurs some hardware modifications. That leaves certain mid-range devices (e.g., ARM Cortex-A family) equipped with standard hardware components, e.g., a memory management unit (MMU) and perhaps a secure boot facility. In this space, seL4 (a verified microkernel with guaranteed process isolation) is a promising platform for attaining RA. HYDRA made a first step towards this, albeit without achieving any verifiability or provable guarantees. This paper picks up where HYDRA left off by constructing a PARseL architecture, that separates all user-dependent components from the TCB. This leads to much stronger isolation guarantees, based on seL4 alone, and facilitates formal verification. In PARseL, We use formal verification to obtain several security properties for the isolated RA TCB, including: memory safety, functional correctness, and secret independence. We implement PARseL in F* and specify/prove expected properties using Hoare logic. Next, we automatically translate the F* implementation to C using KaRaMeL, which preserves verified properties of PARseL C implementation (atop seL4). Finally, we instantiate and evaluate PARseL on a commodity platform -- a SabreLite embedded device.

Abstract: The Internet of Things (IoT) market is rapidly growing and is expected to double from 2020 to 2025. The increasing use of IoT devices, particularly in smart homes, raises crucial concerns about user privacy and security as these devices often handle sensitive and critical information. Inadequate security designs and implementations by IoT vendors can lead to significant vulnerabilities. To address these IoT device vulnerabilities, institutions, and organizations have published IoT security best practices (BPs) to guide manufacturers in ensuring the security of their products. However, there is currently no standardized approach for evaluating the effectiveness of individual BP recommendations. This leads to manufacturers investing effort in implementing less effective BPs while potentially neglecting measures with greater impact. In this paper, we propose a methodology for evaluating the security impact of IoT BPs and ranking them based on their effectiveness in protecting against security threats. Our approach involves translating identified BPs into concrete test cases that can be applied to real-world IoT devices to assess their effectiveness in mitigating vulnerabilities. We applied this methodology to evaluate the security impact of nine commodity IoT products, discovering 18 vulnerabilities. By empirically assessing the actual impact of BPs on device security, IoT designers and implementers can prioritize their security investments more effectively, improving security outcomes and optimizing limited security budgets.

Abstract: Large Language Models (LLMs) have gained widespread popularity across diverse domains involving text generation, summarization, and various natural language processing tasks. Despite their inherent limitations, LLM-based designs have shown promising capabilities in planning and navigating open-world scenarios. This paper introduces a novel application of pre-trained LLMs as agents within cybersecurity network environments, focusing on their utility for sequential decision-making processes. We present an approach wherein pre-trained LLMs are leveraged as attacking agents in two reinforcement learning environments. Our proposed agents demonstrate similar or better performance against state-of-the-art agents trained for thousands of episodes in most scenarios and configurations. In addition, the best LLM agents perform similarly to human testers of the environment without any additional training process. This design highlights the potential of LLMs to efficiently address complex decision-making tasks within cybersecurity. Furthermore, we introduce a new network security environment named NetSecGame. The environment is designed to eventually support complex multi-agent scenarios within the network security domain. The proposed environment mimics real network attacks and is designed to be highly modular and adaptable for various scenarios.

Abstract: In large-scale data analysis, near-duplicates are often a problem. For example, with two near-duplicate phishing emails, a difference in the salutation (Mr versus Ms) is not essential, but whether it is bank A or B is important. The state-of-the-art in near-duplicate detection is a black box approach (MinHash), so one only knows that emails are near-duplicates, but not why. We present DarkDiff, which can efficiently detect near-duplicates while providing the reason why there is a near-duplicate. We have developed DarkDiff to detect near-duplicates of homepages on the Darkweb. DarkDiff works well on those pages because they resemble the clear web of the past.

Abstract: AI programs, built using large language models, make it possible to automatically create phishing emails based on a few data points about a user. They stand in contrast to traditional phishing emails that hackers manually design using general rules gleaned from experience. The V-Triad is an advanced set of rules for manually designing phishing emails to exploit our cognitive heuristics and biases. In this study, we compare the performance of phishing emails created automatically by GPT-4 and manually using the V-Triad. We also combine GPT-4 with the V-Triad to assess their combined potential. A fourth group, exposed to generic phishing emails, was our control group. We utilized a factorial approach, sending emails to 112 randomly selected participants recruited for the study. The control group emails received a click-through rate between 19-28%, the GPT-generated emails 30-44%, emails generated by the V-Triad 69-79%, and emails generated by GPT and the V-Triad 43-81%. Each participant was asked to explain for why they pressed or did not press a link in the email. These answers often contradict each other, highlighting the need for personalized content. The cues that make one person avoid phishing emails make another person fall for them. Next, we used four popular large language models (GPT, Claude, PaLM, and LLaMA) to detect the intention of phishing emails and compare the results to human detection. The language models demonstrated a strong ability to detect malicious intent, even in non-obvious phishing emails. They sometimes surpassed human detection, although often being slightly less accurate than humans.