Large Language Models (LLMs) exhibit impressive performance across various tasks, but deploying them for inference poses challenges. Their high resource demands often necessitate complex, costly multi-GPU pipelines, or the use of smaller, less capable models. While quantization offers a promising solution utilizing lower precision for model storage, existing methods frequently experience significant performance drops at lower precision levels. Additionally, they typically provide only a limited set of solutions at specific bit levels, many of which are extensively manually tuned. To address these challenges, we propose a new method called SKIM: Scaled K-means clustering wIth Mixed precision. Our approach introduces two novel techniques: 1. A greedy algorithm to solve approximately optimal bit allocation across weight channels, and 2. A trainable scaling vector for non-differentiable K-means clustering. These techniques substantially improve performance and can be adapted to any given bit. Notably, in terms of model perplexity, our method narrows the gap between 3-bit quantized LLaMA models and their full precision counterparts by 16.3% on average.

Abstract--As the primary standard protocol for modern cars, the Controller Area Network (CAN) is a critical research target for automotive cybersecurity threats and autonomous applications. The Controller Area Network OBD-II diagnostic data is easy to access via the OBD-II port, (CAN) protocol was firstly developed by Bosch in the as all modern cars are equipped with the OBD-II diagnostic 1980s [1] and serves as the de facto standard protocol for connecting system. OBD-II diagnostic data can be converted into humanreadable ECUs embedded in cars [3]-[5]. The standard structure accurate vehicle data with public formulas to be used of the CAN frame is composed of the start of frame, arbitration in the matching process for associating semantic meanings field, control field, data field, CRC field, ACK field and end with CAN signals. Both OBD-II diagnostic data and regular of frame, as shown in Figure 1. While the CAN protocol has CAN frames can be collected from the OBD-II port. The a standardized frame structure, understanding the protocol's RE systems can leverage both CAN and OBD-II diagnostic utilization for signal transmission remains challenging. This data to create a comprehensive dataset for reverse engineering is because Original Equipment Manufacturers (OEMs) encode purposes, eliminating the need for additional measurement the signals within the CAN frames' data fields (data payloads) equipment like IMUs. in proprietary ways that vary among OEMs, vehicle models, The primary objective of a CAN RE system is to identify the and years [6]. CAN messages frames is the first step to extracting the essential information are structured into frames, and the CAN frames of different to develop autonomous applications or explore automotive CAN IDs have different lengths of the data payload.

Many crypto protocols rely heavily on the security of keys that are stored in device memory and are susceptible to malware attacks. Physically unclonable functions (PUFs) have been proposed as an alternative [1] whose response to external stimuli (challenge) is determined by their microscopic structure, which is difficult to clone. PUF operation consists of two phases: enrollment and deployment. During the enrollment process, the manufacturer creates a challenge response pair (CRP) library by probing the device with unique binary challenges and measuring/generating the corresponding digitized response. The CRP data set is then stored for the deployment phase where the PUF device can be authenticated by probing it with a subset of challenges in the CRP data set and comparing the responses. To be a strong security primitive a PUF must exhibit behavior that is i) deterministic, ii) unpredictable, and iii) unique.