AlKadhim Journal for Computer Science

Pepper Leaf Disease Detection Using Deep Learning Techniques

2025-08-09T13:15:00+03:00

Early diagnosis of pests and plant diseases is crucial for preventing significant crop losses. This study proposes a leaf disease detection system using MobileNetV2 integrated with multiple optimization techniques (Adam and learning rate scheduling). Evaluated on the Pepper PlantVillage dataset, the MobileNetV2 model employs patch embedding and attention mechanisms for feature extraction, with SoftMax used for final classification. The model was further validated on a multi-class Apple PlantVillage dataset. Results demonstrate high accuracy: 97.03% for pepper and 94.63% for apple classification. Comparative analysis with CNN architectures shows superior efficiency and faster convergence for our model, outperforming Inception v3 (96.81%) and VGG-19 (94.93%) on the pepper dataset.

AI- Techniques for Decoding Language Representation from EEG-Based Brain Activity: A Comprehensive Review

2025-07-09T21:43:54+03:00

This comprehensive paper examines artificial intelligence techniques for decoding and representing language from brain activity, as measured by electroencephalography (EEG), a common, non-invasive, and high-temporal-resolution method for recording neural activity from the brain during language processing. This technique faces several challenges in dealing with noise and complexity. It also reviews the significant progress in analyzing this signal using traditional machine learning techniques such as SVM, RF, as well as deep learning techniques like CNN, LSTM, and Transformers. This paper also discusses the main applications of these technologies in providing communication tools for people with disabilities, medical diagnosis and treatment, understanding linguistic perception, as well as the challenges related to data quality, cost, complexity, and ethical issues. It also offers promising future insights into the integration of multiple technologies, predicting neurological and cognitive conditions, and developing advanced brain-computer interfaces, paving the way for a deeper understanding of language processing mechanisms in the brain.

Cognitive Honeypots AI-Enhanced Deception for Proactive Threat Hunting

2025-09-22T08:12:41+03:00

The rapidly increasing complexity of cyber threats, including AI-powered attacks, is forcing a shift in defense strategies from reactive to proactive approaches. Traditional honeypots remain largely static and easily identifiable, while newer AI-enhanced models emphasize realism but still lack deep understanding of attacker cognition. To address this gap, this paper introduces CogniTrap, a novel framework that combines a high-interaction honeypot with an AI-driven cognitive deception engine. CogniTrap dynamically creates and adapts “cognitive decoys” designed to exploit attackers’ biases and reasoning flaws. A prototype of CogniTrap was developed and deployed, where decoy placements and adaptations were optimized using reinforcement learning informed by live analysis of attacker tactics, techniques, and procedures (TTPs). Intelligence gathered from triggered decoys was transformed into proactive hypotheses for threat hunting in production environments. Experimental results from comparative 30-day live deployments showed that CogniTrap increased attacker dwell time by 45% compared to a standard high-interaction honeypot and generated higher interaction rates with deceptive assets. Furthermore, it was able to produce high-fidelity threat hunting queries based on attacker cognitive patterns, validating its practical utility. This research marks the first implementation-based framework for adaptive cognitive honeypots, bridging the gap between theoretical cognitive security concepts and operational proactive threat hunting. By providing architecture, algorithms, and empirical validation, CogniTrap establishes a new paradigm for intelligent cyber defense.

Audio Encryption and Decryption using the Affine Cipher with the One-Time Pad (OTP)

2025-08-07T16:57:15+03:00

This research aims to design and implement a hybrid encryption system to secure audio signals, based on the integration of two cryptographic algorithms: Affine Cipher and One-Time Pad (OTP). The Affine Cipher algorithm is a simple and efficient model of classical encryption, while the One-Time Pad algorithm provides a high level of security when using a random, unrepeated key. The proposed system was applied to a digital audio signal, where the signal was first converted to a numerical representation (8-bit), then encrypted using Affine Cipher, followed by the application of OTP encryption to the resulting outputs. The practical evaluation included the use of quantitative measures such as the average quadratic error (MSE), the average absolute error (MAE), and the signal-to-noise ratio (SNR), in addition to the statistical distribution of signals at each stage of the encryption by histogram. The results showed that the loose signal was exactly identical to the original signal (MSE = 0, MAE = 0, SNR = ∞), and statistical analysis of the histogram proved the effectiveness of the system in blurring the frequency patterns of the signal, which enhances its resistance to analytical attacks. The proposed system provides a balance between simplicity of implementation and efficiency in encryption, and is suitable for securing audio data in real-time applications and sensitive systems

Multi-Dimensional Deep Learning for Unusual Detection in Security Footage: A 3D CNN, LSTM, and Attention-Based Approach

2025-08-15T09:23:40+03:00

Automated detection of unusual activities in surveillance videos remains a critical challenge due to the vast volume of footage and the rarity of anomalous events. This study proposes a novel deep learning framework that integrates 3D convolutional neural networks for spatiotemporal feature extraction, Long Short-Term Memory networks for modeling temporal dependencies, and an attention mechanism to focus on salient segments. The primary objective is to achieve highly accurate binary classification of video clips into “usual” and “unusual” categories, while addressing class imbalance and environmental variability. The model is trained and evaluated on three large-scale datasets, UCF‑Crime, XD-Violence, and CCTVFights, which comprise surveillance videos covering real-world anomalies. Experimental results show that the proposed method achieves an overall accuracy of 97.41% on the UCF-Crime dataset, 98.11% on the XD-Violence dataset, and 98.50% on the CCTVFights dataset, in addition to high precision, recall, and F1 scores on the three datasets used in the evaluation process, outperforming existing benchmarks. These findings indicate that combining spatiotemporal modeling and attention-driven context aggregation can significantly enhance anomaly detection performance in complex surveillance scenarios.