Junaid Ahmad Khan

Post-training 4-bit quantization is often treated as the default path for running large language models (LLMs) on prosumer GPUs: if decoding is memory-bandwidth bound, shrinking weights from FP16 to 4-bit should cut memory traffic and improve latency and energy efficiency. We revisit this assumption on an RTX 3090 (Ampere) that lacks native INT4 tensor support. For 1-8 billion parameter models (TinyLlama, Qwen-2.5, Mistral, Llama 3.1, DeepSeek-R1-8B), we compare native FP16 inference against AutoGPTQ 4-bit models and GGUF kernels in llama.cpp. On a standard Transformers+AutoGPTQ stack, interactive batch size B = 1 decoding is still 1.3-2.2× slower than FP16 despite a 2.4× reduction in VRAM usage, and some INT4 configurations are up to 2.4× less energy-efficient. An optimized GGUF backend improves 4-bit TinyLlama throughput by 1.65× over GPTQ, indicating that the de-quantization penalty is dominated by kernel design rather than hardware limits. We conclude that on prosumer GPUs without native INT4 tensor cores, 4-bit quantization is only attractive when paired with mature low-bit kernels; otherwise, FP16 remains the more robust choice for interactive workloads.

Post-training 4-bit quantization is often treated as the default path for running large language models (LLMs) on prosumer GPUs: if decoding is memory-bandwidth bound, shrinking weights from FP16 to 4-bit should cut memory traffic and improve latency and energy efficiency. We revisit this assumption on an RTX 3090 (Ampere) that lacks native INT4 tensor support. For 1–8 billion parameter models (TinyLlama, Qwen-2.5, Mistral, Llama 3.1, DeepSeek-R1-8B), we compare native FP16 inference against AutoGPTQ 4-bit models and GGUF kernels in llama.cpp. On a standard Transformers+AutoGPTQ stack, interactive batch size $B=1$ decoding is still 1.3–2.2× slower than FP16 despite a 2.4× reduction in VRAM usage, and some INT4 configurations are up to 2.4× less energy-efficient. An optimized GGUF backend improves 4-bit TinyLlama throughput by 1.65× over GPTQ, indicating that the de-quantization penalty is dominated by kernel design rather than hardware limits. We conclude that on prosumer GPUs without native INT4 tensor cores, 4-bit quantization is only attractive when paired with mature low-bit kernels; otherwise, FP16 remains the more robust choice for interactive workloads.

In-vehicle networks (IVNs) connect numerous sensors, processors, and Electronic Control Units (ECUs) to enable real-time communication, but their growing complexity has amplified cybersecurity concerns. The Controller Area Network (CAN) bus, along with its variants CAN-FD, CAN-XL, and complementary protocols such as LIN and FlexRay, remains the backbone of IVNs. However, these protocols (the CAN specifically) were not originally designed to meet the stringent security demands of emerging autonomous and electric vehicles. This survey critically examines existing IVN security literature with a particular focus on intrusion detection systems (IDS) developed for CAN-based architectures. Unlike prior reviews that emphasize algorithmic design or attack taxonomy, this survey provides a comparison of IDS approaches (typically CAN-based due to its dominance in industry and research domain) based on physical-layer features, statistical parameters, information theory, and learning-based methods. We highlight unresolved security limitations of the CAN bus, synthesize mitigation strategies across industrial domains (automotive, aerospace, marine, agricultural, and battery systems), and identify overlooked challenges. Finally, we outline key research frontiers, such as hardware-assisted protection, edge intelligence, blockchainbased authentication, generative AI-driven anomaly detection, and quantum-safe cryptography—toward building resilient next-generation IVNs.

Precision agriculture demands robust yet efficient models capable of operating on resource-constrained edge devices for real-time plant health monitoring. Existing Vision Transformer (ViT) models often underperform in data-scarce agricultural settings due to their reliance on large-scale pretraining and limited local feature extraction capabilities. In this study, we propose a enhanced lightweightViT architecture (EViT) for edge devices used in plants disease classification. The core of our approach is a domain-optimized convolutional stem (ConvStem) architecture that replaces the standard patch embedding with a convolutional stem to enhance local feature extraction, particularly effective in data-scarce scenarios. Unlike ViT architectures, which rely on large-scale pretraining, the proposed ConvStem-ViT (i.e., EViT) achieves a notable improvement in accuracy and generalization on two benchmark datasets - PlantVillage (a widely used dataset of leaf disease images) and CCMT (a more challenging, real-world crop disease dataset). Comprehensive experiments, including ablation studies using model variant analysis, supported with attention map visualizations, confirm the superiority of the proposed model over standard ViTs and contemporary CNN-based baselines. Our EViT achieves over 94.8% accuracy on PlantVillage and 78.0% on CCMT, outperforming conventional ViTs by up to 13.6%. Computational analysis shows the model achieves these gains with minimal overhead (5.55 ms/image latency, 1.30 GFLOPs); these efficiency metrics enable real-time inference on UAVs or low-power edge devices, making the model suitable for practical drone-based scouting in agricultural fields.. This work offers a viable framework for scalable, accurate, and efficient plant disease classification, making it an effective choice for resource-constrained edge devices such as drones for plant disease classification.

Driver behavior features extracted from the controller area network (CAN) have potential applications in improving vehicle safety. However, the development of a classifier-based IDS for in-vehicle networks remains an open research problem. To address this challenge, we incorporate novel n -fold cross-validation windowing techniques on two publicly available driving behavior datasets. A driver classification-based IDS is proposed using the LSTM-FCN model that utilizes the strengths of both fully convolutional network (FCN) and long short-term memory (LSTM) networks. These modules allow the model to learn spatial and temporal features and utilize contextual information. In addition, we combine three squeeze and excite (SnE) layers following FCN layers to incorporate adjacent spatial locations and augment a scaled dot product attention mechanism into the LSTM to improve its feature selection and extraction capabilities. Our proposed IDS uses hacking and countermeasure research lab (HCRL) and test datasets, which achieve an improvement in accuracy of 4.18% and 13.99% respectively, from the baseline LSTM-FCN model. The experimental results of our method exhibited an overall accuracy of 99.36% and 96.36% for both datasets and outperformed various state-of-the-art methods.

For in-vehicle network communication, the controller area network (CAN) broadcasts to all connected nodes without address validation. Therefore, it is highly vulnerable to all sorts of attack scenarios. This research proposes a novel intrusion detection system (IDS) for CAN to identify in-vehicle network anomalies. The statistical characteristics of attacks provide valuable information about the inherent intrusion patterns and behaviors. We employed two real-world attack scenarios from publicly available datasets to record a real-time response against intrusions with increased precision for in-vehicle network environments. Our proposed IDS can exploit malicious patterns by calculating thresholds and using the statistical properties of attacks, making attack detection more efficient. The optimized threshold value is calculated using brute-force optimization for various window sizes to minimize the total error. The reference values of normality require a few legitimate data frames for effective intrusion detection. The experimental findings validate that our suggested method can efficiently detect fuzzy, merge, and denial-of-service (DoS) attacks with low false-positive rates. It is also demonstrated that the total error decreases with an increasing attack rate for varying window sizes. The results indicate that our proposed IDS minimizes the misclassification rate and is hence better suited for in-vehicle networks.