What question did this study set out to answer?

The research aims to develop a reliable method for detecting deepfakes in multimedia communications using advanced machine learning techniques.

March 19, 2026Open Access

Integrating enhanced feature fusion with adaptive hidden semi-markov model and deep neural network for robust deepfake detection in Internet of Things-based edge computing environments

Key Points

The research aims to develop a reliable method for detecting deepfakes in multimedia communications using advanced machine learning techniques.
Utilized data augmentation, image resizing, and normalization to improve image quality.
Integrated feature fusion from ResNet-50, EfficientNet-B4, and MobileNetV3 for feature representation.
Employed adaptive hidden semi-Markov model to capture temporal dependencies in data.
Implemented a deep belief network for effective classification of real and fake content.
Achieved an impressive detection accuracy of 98.61% using the Celeb DF (v2) dataset.
Demonstrated improved generalization for distinguishing real from fake multimedia content.

Abstract

The increasing convergence of Internet of Things (IoT) devices and edge computing platforms in current digital networks has presented extraordinary opportunities. However, this surge has presented a massive attack surface susceptible to developing attacks like deepfakes. Deepfakes refer to realistic, but fake images, videos, and sounds made by artificial intelligence (AI) models. Recent developments in deepfake generation have made deepfakes more accurate and simpler to make. Deepfakes have previously been employed for numerous malicious reasons, including the publication of fabricated outcomes in scientific magazines or an attack on the identity tests employed by banks over artificial videos and voices, raising concerns regarding them and their usage. In response to this occurrence, the research group has selected the growth of techniques to distinguish real content from deepfakes. Machine learning (DL) and deep learning (DL) driven systems have attained extremely precise outcomes in the detection of deepfakes. Furthermore, numerous previous studies on perceiving deepfake images or videos were concentrated on DL. This paper presents a Feature Fusion Integrated with Adaptive Hidden Semi-Markov Model for Robust Deepfake Detection (FFAHSMM-RDD) method. The aim is to present a reliable pathway for securing multimedia communication and preventing deepfake-driven threats in real-world applications. Initially, data augmentation, image resizing, and normalization techniques are applied to improve image quality and enhance generalization. For feature representation, a fusion of three models, such as ResNet-50, EfficientNet-B4, and MobileNetV3, is used. Moreover, the adaptive hidden semi-Markov model (AHSMM) is employed to model temporal dependencies and sequence dynamics within the data effectively. In the deepfake classification process, a deep belief network (DBN) model is implemented to classify real and fake content efficiently. The performance validation of the FFAHSMM-RDD methodology demonstrated a superior value of 98.61% under the Celeb DF (v2) dataset.

KI fragen

Bookmark

View Full Paper