In digital forensic investigations, content-based image retrieval (CBIR) is essential for identifying visually similar evidence, such as suspect photographs, crime scene captures, or illicit digital content, from large-scale image repositories. The sensitive nature of forensic data demands stringent privacy protection, especially when storage and processing are delegated to untrusted cloud environments. Traditional encryption methods safeguard confidentiality but prevent similarity search without decryption, posing a significant barrier to efficient investigations. To overcome this limitation, we propose a privacy-preserving CBIR framework for digital forensics that integrates deep learning-based feature extraction with homomorphic encryption. First, a convolutional neural network (CNN) model is employed to derive high-dimensional, discriminative descriptors from forensic images. These descriptors are then transformed into the encrypted domain, enabling similarity computations to be performed directly on ciphertexts without exposing the original image content. Our approach ensures both retrieval effectiveness and strict data confidentiality, resisting leakage and inference attacks while complying with legal and operational requirements. Experiments on benchmark and synthetic forensic datasets show that the proposed system maintains high retrieval accuracy and acceptable computational overhead, making it practical for real-world law enforcement workflows. This work advances secure, privacy-aware multimedia search in the context of digital forensics, balancing investigative utility with robust privacy guarantees.