Document Type : Original Research Paper
Author
PhD Student, Department of Computer Engineering, Faculty of Technical and Engineering, Islamic Azad University, South Tehran Branch, Tehran, Iran.
Abstract
BACKGROUND AND OBJECTIVES: One of the major challenges in the insurance industry is combating fraud during the claim submission process, which is particularly evident in digital claim systems where customers send images of damages and injuries through mobile applications. As the insurance sector continues its digital transformation, introducing modern methods across various operational stages such as underwriting, issuance, and claim processing, the necessity of developing innovative and secure fraud detection mechanisms has become more crucial than ever. With the rapid digitization and digitalization of insurance services, ensuring the authenticity of transmitted data and protecting sensitive information have emerged as key priorities. In digital claim reporting, where policyholders upload damage photographs via mobile applications, the authenticity of spatial and temporal metadata—specifically the location and time of the captured images—plays a vital role in reducing fraudulent submissions. To address this issue, this study proposes a novel security framework that utilizes steganography and cryptography for embedding encrypted metadata directly into the image file. In this method, the location and timestamp information extracted from the damage images are first encrypted using the AES-GCM algorithm and then securely embedded into the image through a Generative Adversarial Network (GAN)-based steganographic model. The hidden data can subsequently be extracted and decrypted at the insurer’s end for verification. This approach not only strengthens the authenticity verification of the submitted images but also enhances data confidentiality and resistance to tampering. By integrating this method into a company’s insurance application, insurers can build greater trust with policyholders, automate the verification process, and effectively reduce costs associated with fraudulent activities. The combination of deep learning–based steganography with robust encryption offers a promising path toward intelligent, secure, and fraud-resistant insurance claim systems.
METHODS: The proposed system employs a GAN-based steganographic framework to embed spatial and temporal metadata within damage images. Initially, the Exif metadata (including the capture time and GPS coordinates) is extracted and encrypted using the AES-GCM (Advanced Encryption Standard – Galois/Counter Mode) algorithm. This encryption stage ensures confidentiality and integrity of sensitive data before it is hidden in the image. The encrypted data are then embedded using a deep neural network generator, which learns to conceal information within the least perceptible image regions while preserving high visual fidelity. For performance evaluation, the system was tested on a dataset of 100 automobile damage photographs having been collected from real insurance claim records. Three methods were compared: Baseline LSB (Least Significant Bit) method, GAN-based steganography without encryption, and GAN combined with AES-GCM encryption (proposed method). The evaluation considered several quantitative metrics: Image quality measured by Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM), Embedding capacity, Bit Error Rate (BER) under standard steganalysis attacks such as JPEG compression (quality 50), Gaussian noise (σ = 5), and image scaling (0.9), and Detection rate of metadata manipulation to assess system robustness against tampering.
FINDINGS: The experimental results confirmed the effectiveness of the proposed approach. Images containing hidden data maintained a high visual quality with an average PSNR of 41 dB, demonstrating that the steganographic embedding had minimal perceptual impact. The GAN-based method outperformed the baseline LSB technique, achieving higher image quality (40.6 ± 0.8 dB vs. 36.7 ± 1.1 dB) and better structural similarity. Moreover, the average BER under compression and noise attacks was nearly half that of the baseline method (≈12–19% vs. ≈29–34%). Integrating AES-GCM encryption had no negative effect on visual quality but significantly improved end-to-end data security. Statistical analysis using a paired-sample t-test revealed that the performance differences were significant at a 95% confidence level (p < 0.001). The 95% confidence interval for PSNR values of the proposed model was calculated to be [40.7, 41.3], reflecting the system’s stability and consistency. In the metadata tampering experiments, all instances of manipulated metadata in the 100-image dataset were successfully detected, achieving 100% detection accuracy. The average mobile implementation time for the complete encryption and embedding sequence was approximately 2.0 seconds per image, confirming the feasibility of applying the system in real-time insurance applications
CONCLUSION: The findings of this research demonstrate that employing deep learning-based steganography (GAN) in combination with AES-GCM encryption can effectively mitigate security challenges in online insurance claim processes. The proposed method provides a reliable mechanism to verify the authenticity of spatial and temporal data while maintaining high image quality and computational efficiency. This integrated framework enhances both data protection and fraud prevention, ensuring that insurance companies can verify the legitimacy of submitted claim images with greater confidence. Consequently, the proposed GAN–AES hybrid model achieves an optimal balance between image fidelity, robustness to steganographic attacks, and metadata integrity verification. The results suggest that this approach can serve as a practical and efficient layer of protection for digital insurance claim systems, contributing to enhancing trust, transparency, and security across the insurance ecosystem.
Keywords
Main Subjects
)
Letters to Editor
Send comment about this article