The main objective of this project is to identify overspeed vehicles, using Deep Learning and Machine Learning Algorithms. After acquisition of series of images from the video, trucks are detected using Haar Cascade Classifier. The model for the classifier is trained using lots of positive and negative images to make an XML file. This is followed by tracking down the vehicles and estimating their speeds with the help of their respective locations, ppm (pixels per meter) and fps (frames per second). Now, the cropped images of the identified trucks are sent for License Plate detection. The CCA (Connected Component Analysis) assists in Number Plate detection and Characters Segmentation. The SVC model is trained using characters images (20X20) and to increase the accuracy, 4 cross fold validation (Machine Learning) is also done. This model aids in recognizing the segmented characters. After recognition, the calculated speed of the trucks is fed into an excel sheet along with their license plate numbers. These trucks are also assigned some IDs to generate a systematized database.

Modern and future vehicles are complex cyber-physical sys-tems. The connection to their outside environment raises many securityproblems that impact our safety directly. In this work, we propose a DeepCAN intrusion detection system framework. We propose a multivariatetime series representation for asynchronous CAN data. This represen-tation enhances the temporal modelling of deep learning architecturesfor anomaly detection. We study different deep learning tasks (super-vised/unsupervised) and compare different architectures, to propose anin-vehicle intrusion detection system that fits constraints of memory andcomputational power of the in-vehicle system. The proposed intrusiondetection system is time window wise: any given time frame is labelledeither anomalous or normal. We conduct experiments with many types ofattacks on an in-vehicle CAN using SynCAn dataset. We show that oursystem yields good results and allow to detect different kinds of attacks.

AEGIS-Omega is a high-performance, hybrid multi-layer Intrusion Detection System (IDS). It features a 4-layer detection strategy—Signature Analysis, Autoencoder Anomaly Detection, BiLSTM Deep Learning, and Ensemble Fusion—to identify attack types with 95% F1-score. Includes a FastAPI backend, React dashboard, and PCAP,Netflow analysis.

In this project, I developed a mullitvariate time-series deep learning model (LSTM) for prediction of energy use (kWh) in residential buildings. The dataset used for this experiement is a daily dataset collected for two different house IDs (i.e. 624 and 1103) for the year 2015. Each house ID has 365 datapoints with the following feature variables: id, timestamp, actual_energyuse (kWh), avg_temp_c (celcius). We evaluated the LSTM model by computing the test error (RMSE) and the coefficient of variation of RMSE and the following results were obtained for both house IDs: house id 624 (Test RMSE: 8.83, CV-RMSE: 1.54) , house id 1103 ( Test RMSE: 4.49, CV-RMSE: 0.93)

Deep Learning IDS 4 Low Power and Lossy Networks

A machine-learning-based IDS for real-time network threat detection using signature & anomaly-based methods. Built with Python, Scikit-learn, it processes the UNSW_NB15 dataset, achieving 95.6% accuracy, 96.4% recall with Random Forest + RFE. Future work includes deep learning, hyperparameter tuning & real-world deployment.

This project develops scalable IDS using distributed deep learning on the BoT-IoT dataset. It evaluates centralized multi-node (Docker), centralized multi-GPU, and decentralized multi-GPU setups. A CNN with SE blocks achieved >98.9% accuracy on a 4-GPU (NVIDIA RTX 5000), 20-core Intel i9 high-performance CPU.