ABSTRACT. Hate speech is racing across social networks, choking the online space with danger, as bitter smoke fills a packed room. Traditional text models often overlook subtle hate, especially when it sneaks in as a sarcastic jab, a clipped, biting tone, or a meme whose grin cuts deeper than the caption. This study introduces a multimodal deep learning framework that brings together text, emotion, sarcasm, and visual signals, such as a quick smile or a sharp tone to boost detection accuracy. The model blends DeBERTa (Decoding-enhanced BERT with Disentangled Attention)’s nuanced understanding of text with DistilBERT (Distilled Bidirectional Encoder Representations from Transformers)’s ability to capture emotional and sarcastic cues, and CNN (Convolutional Neural Network)’s sharp eye for visual details enhanced by OCR (Optical Character Recognition) for embedded text, weaving them together to spot hateful content with precision. The tests revealed that the method beats single-modality baselines, catching hateful language with sharper accuracy and a better grasp of context, such as detecting the sting in an offhand remark.

ABSTRACT. Pneumonia remains a serious global health challenge, demanding quick, precise diagnosis from chest X-rays yet manual reading often drags on and can vary wildly from one doctor to another, like trying to judge a faint shadow on a dim film. This project rolls out an end-to-end, explainable AI that spots whether a chest X-ray say one with a faint shadow near the lung is normal or shows signs of pneumonia. We use a sturdy dual model setup built around two Convolutional Neural Networks ResNet-18 and DenseNet-121 both pretrained on ImageNet, sharp and ready like tools fresh from the workshop. Blending these models into a weighted ensemble sharpens predictions and keeps results steady, like a compass holding true in rough weather.We employ Gradient-weighted Class Activation Mapping (Grad-CAM) to produce dynamic heatmaps that identify the precise areas influencing the model's choice, illuminating the most significant features with a warm glow in order to maintain the transparency of our process. Even when you're exhausted and your coffee has gone cold, the system's interface remains user-friendly and it operates on a minimalist Flask web app with a confidence level that reduces false positives. In order to provide doctors with a decision-support tool that is both effective and simple to use like reading clear text under a bright lamp rather than squinting at a smudge this study blends cutting-edge deep learning with essential interpretability.

ABSTRACT. Parkinson’s disease moves in quietly, progressing over time, so catching it early without invasive tests is crucial.This study presents a detailed framework that uses subtle changes in voice like a faint pitch tremor or a slight shimmer to help iden tify Parkinson’s disease. We compare seven models ranging from classic machine learning classifiers to a 1D Convolutional Neural Network using the standardized Parkinson’s Voice Dataset, where each voice clip carries the faint tremor of the condition. Our results show that ensemble methods work best for diagnosing voices, and the XGBoost classifier topped the charts with a 94.87 percentage accuracy rate. Alongside the voice module, the system packs a second unit built to analyze spiral and wave images,tracing their curves as if following ripples in water. This module uses a pre-trained CNN and applies Grad-CAM (Gradient weighted Class Activation Mapping) to make its predictions easier to understand, lighting up the exact parts of a drawing like a dark corner or a curved line that steer its decision. The study confirms that voice features can predict outcomes with striking accuracy, sets a high bar for performance, and shows that an integrated, multimodal, explainable AI system could realistically guide future clinical decisionsmuch like a clear, confident voice cutting through background noise.

ABSTRACT. Diabetic Retinopathy (DR) remains a leading cause of preventable blindness worldwide, affecting approximately one-third of diabetic patients. Early detection through retinal fundus imaging is crucial for preventing vision loss, but manual screening presents challenges in scalability and inter-observer variability. This paper presents a novel hybrid deep learning architecture that integrates ResNet50 for hierarchical feature extraction with Capsule Networks for precise spatial relationship modeling. The proposed model addresses critical limitations of conventional Convolutional Neural Networks (CNNs), particularly their inability to preserve spatial hierarchies and part-whole relationships essential for accurate DR severity grading. Our approach leverages ResNet50's powerful feature extraction capabilities to identify intricate retinal patterns, while Capsule Networks capture spatial dependencies through dynamic routing and vector outputs. Experimental results on retinal fundus images demonstrate superior performance compared to standalone CNN architectures, with significant improvements in detecting subtle pathological features such as microaneurysms, exudates, and hemorrhages. The hybrid model achieved an overall accuracy of 97.2 in five-class DR classification, showcasing strong potential as an automated screening tool for clinical deployment.

ABSTRACT. Aging is among the most critical phenomena that affect the reliability of integrated circuits (ICs), because of which the parameters of the circuit worsen with time and can even lead to the circuit failure. In this paper, we present a comprehensive analysis of the impact of supply voltage scaling on aging and delay characteristics of CMOS INVERTER, AND, OR gates and propose using high power supply VDD levels to minimize aging effects.

ABSTRACT. The article presents a new technique for improving the PSRR in voltage regulator with a PMOS output transistor. The voltage regulator with a PMOS output transistor was designed to support a 2-mA active load. The technique is relevant for modern 14nm and below technology nodes. SPICE analysis performed over the designed circuit proved the worst case PSRR value as of -35 dB at lower frequencies at wide temperature ranges of -40..125 ° C and supply voltages up to 0.994 V considering the process deviation within ± 4.5 Sigma. Voltage overshoot protection circuit was also developed to ensure safe operation of core devices in various operation modes.

ABSTRACT. The design and fabrication of integrated circuits (ICs) increasingly rely on outsourcing, involving multiple third-party entities. This multi-tiered supply chain introduces vulnerabilities, enabling adversarial actors to implant Hardware Trojans (HTs) at various stages of development. Such malicious modifications can lead to a wide range of security threats, including information leakage (e.g. MOLES Trojan) and denial-of-service (rarely triggered bit flip). Although various methods exist for generating tests to detect HTs, they face significant challenges due to the immense complexity of the search space. This complexity renders the process impractical and results in insufficient trigger coverage. Effective HT detection depends on using suitable test vectors. This paper introduces a novel algorithm, termed GUARD, which synergistically combines the Artificial Bee Colony (ABC) algorithm and Genetic Algorithm (GA), along with a machine learning-based clustering technique for dimensionality reduction, within an automated framework for HT detection. Experimental evaluations on the ISCAS-85 and ISCAS-89 benchmark suites demonstrate a substantial increase in stimulus coverage and a marked reduction in execution time compared to existing state-of-the-art techniques.

ABSTRACT. modern high-speed SerDes and clock/data recovery (CDR) systems heavily rely on phase interpolators (PI) to generate accurate clocks for timing critical paths. As data rates continue to scale, achieving linear code-to-phase transfer becomes increasingly difficult due to device nonidealities, shrinking supply voltages, and tighter jitter budgets. Nonlinearity in the interpolator directly degrades timing accuracy, leading to higher bit-error rates (BER). This work explores methods to improve interpolator linearity in CMOS inverter-based mixer architectures. The three methods are: digital predisposition, an ac-coupled feedback inverter at the input, and reduced supply operation of the mixer stage. Simulation results show that the combination of these methods effectively flattens the code-to-phrase characteristics and reduces integral nonlinearity. Trade-offs are added circuit complexity, area and jitter. The results demonstrate that with modest design, the proposed methods provide a practical way toward more accurate PIs in advanced SerDes and CDR applications.

ABSTRACT. Convolution Neural Network (CNN) is considered now the state of the art of the image processing and classification due to its high accuracy in many applications like autonomous driving and face detection, but in order to achieve this accuracy, this comes with a price of using computationally expensive and high power consumption graphical processing unit (GPU) which is also not always suitable to be used in Real time applications and mobile devices, so coming up with solutions for acceleration and reducing power consumption is a now a need. Field Programmable Gate Array (FPGA) can provide a good alternative due to its low power consumption and dynamic re-configurability. This paper presents a hybrid hardware/software implementation of a faster-RCNN-based object detection system for autonomous driving. Accelerating computations by implementing the VGG-16 backbone network on an FPGA using High-Level Synthesis (HLS) is proposed, aiming to enhance inference speed, efficiency and reduce power consumption by using fixed point calculations instead of floating point. The paper explores the trade-offs between software and hardware implementations, discussing performance improvements and feasibility for real-time applications.