File fragment type identification is a critical research topic in digital forensics. However, current methods still face numerous challenges, including low accuracy, reliance on manually designed features, and limited coverage of file types. To address these limitations, inspired by binocular vision principles, we propose a Parallel Cross Convolutional Neural Network (PCCNN) architecture. This architecture overcomes the limitations of single-structure design and insufficient feature extraction in existing models. The PCCNN can automatically learn features from raw bytes. It extracts features through parallel left and right branch networks, conducts multiple feature concatenation and convolution operations, and fuses feature information from different branches, thereby achieving more comprehensive feature extraction and improving model performance. On the FFT-75 dataset, our model outperforms FiFTy. Compared with the FiFTy model, the average accuracy across six scenarios is improved by 15
Dire consequences, from money and identity theft to the leaking of confidential data, can be caused by cyberattacks. Therefore, deploying an Intrusion Detection System (IDS) to identify such attacks is important. IDS only raises an alarm on the detection of intrusion; however, stopping such attacks before they cause damage is crucial. Analysis of activity logs using Deep Learning (DL) approaches can help gain insights into mitigating such threats. Hence, the proposed work develops a predictive cybersecurity framework for detecting network intruder behaviour and forecasting the next action of an attacker during data transmission. Here, the proposed framework models the sequential attacker behaviour using honeypot interaction logs, enabling accurate and robust detection of intruder behaviours. As an attack depends on the previous actions performed by the intruder, Gated Recurrent Unit (GRU) as well as Long Short-Term Memory (LSTM) are wielded for modelling the system. GRU and LSTM can handle time series data and find patterns in it. The IP address source is utilized to identify the intruder, and a dense vector representation is devised with the help of embedding. These models give an accuracy of approximately 96
Multimodal data analysis plays a pivotal role in understanding human-centric content by integrating text, visual, and audio cues. Humor detection in videos remains particularly challenging due to temporal misalignment across modalities, data imbalance, and noise in real-world signals. To address these challenges, we propose a novel framework for multimodal humor detection that combines Domain-Specific Representation Learning (DSRL) with a Deep Hybrid Fusion Network (DHFN) Our approach disentangles modality-invariant and modality-specific features through a Bi-LSTM based DSRL module and integrates them via cross-modal attention with suppression mechanisms in DHFN. Experiments on benchmark datasets including IEMOCAP, CMU-MOSI, and CMU-MOSEI demonstrate that the proposed model outperforms state-of-the-art baselines, achieving up to 95.8
