Page 183 - tech fest 2025 ב״ש חוברת תקצירים

Deepfake Audio - An AI-Based System for Deepfake Audio Detection
SE-D-09
Eden Edry; edened2@ac.sce.ac.il Noa Rofe; noaro2@ac.sce.ac.il
Advisor: Ms. Alona Kutsyy
SCE - Shamoon College of Engineering, Be’er-Sheva
Deepfake technology poses threats to security, privacy, and trust, with audio manipulation becoming a growing concern. While deepfake detection in images and videos has advanced, identifying synthetic audio remains challenging. Deepfake Audio is an artificial intelligence (AI)-powered platform designed to detect and analyze deepfake audio. The system leverages machine learning (ML) and deep learning (DL) techniques, including spectral analysis and anomaly detection, to distinguish between real and manipulated audio. It processes speech, extracts acoustic features, and applies classification models to verify authenticity. Users benefit from real-time analysis, confidence scores, and an intuitive interface. Deepfake Audio enhances cybersecurity, protects media integrity, and raises awareness of deepfake risks. It also serves as an educational tool to improve digital literacy regarding synthetic media.
Keywords: cybersecurity, deep learning (dl), deepfake audio, education, machine learning (ml), media integrity
Optical Character Recognition (OCR) of Handwritten Hebrew Documents
SE-D-10
Tomer Kakou; tomerkakou@gmail.com Tal Bo-Ahron; talbo4@gmail.com
Advisor: Dr. Natalia Vanetik
SCE - Shamoon College of Engineering, Be’er-Sheva
As digital transformation accelerates, processing handwritten text images is increasingly crucial for searching, storage, and editing. While OCR for printed text has advanced, handwritten Hebrew remains challenging. We are developing a method to integrate a curated dataset with the Hebrew HDD dataset for better classification. Our approach enhances image resolution and extracts character images using OpenCV. Each character is classified into 27 Hebrew letter classes, and then reassembled into words. We use deep learning models like Vision transformer (ViT) and ResNet-50 for recognition, with large language model (LLMs) providing contextual corrections. Evaluation metrics include character error rate (CER), word error rate (WER), and normalized Levenshtein distance (NLD). Our goal is an end-to-end Hebrew handwriting OCR solution with high accuracy, even for degraded text.
Keywords: Hebrew, image classification, low-resource languages, OCR
Book of Abstracts | 2025
183