A Person Re-identification Scheme That Preserves Privacy


Person re-identification (Person Re-ID) is a sophisticated computer vision approach that makes identifying people using surveillance cameras at different places and times easier. Using personal images poses substantial privacy concerns, even though it has a huge potential to improve security and public safety. Since individual images count as private information under data privacy laws and regulations, these problems require privacy-preserving solutions.

Existing approaches to privacy-preserving person Re-ID face certain limitations. Conventional encryption methods provide strong privacy guarantees but fail to allow computations over encrypted data. Homomorphic encryption (HE) directly supports calculations over ciphertexts but doesn’t allow the cloud server to access computation results. Additionally, existing encryption mechanisms for floating-point feature vectors suffer from decoding and calculation errors.

Recently, a new article was published to propose a new privacy-preserving person Re-ID solution called FREED. This system formulates privacy-preserving person Re-ID as similarity metrics of encrypted feature vectors, enabling the cloud server to perform Re-ID operations without compromising any personal image privacy. 

Concretely, FREED utilizes new encoding mechanisms and secure batch computing protocols to encrypt floating-point feature vectors and perform Re-ID operations effectively.

FREED introduces three key components to protect the privacy of feature vectors during the process:

  1. The Encoding Mechanism (ECMO) converts floating-point feature vectors into integers, ensuring accuracy and avoiding decoding errors.
  2. The Secure Batch Multiplication (BatchSMUL) protocol efficiently computes similarity metrics of encrypted feature vectors, reducing computation costs.
  3. The Secure Batch Partial Decryption (BatchPDec) protocol securely ranks the similarity metrics, enabling accurate person re-identification without compromising individual privacy.

Together, these components provide a robust, privacy-preserving solution for person re-identification tasks.

Using ECMO is proposed to convert floating-point feature vectors into integers, which offers two key advantages. Firstly, it eliminates decoding errors commonly encountered in other encoding methods. ECMO ensures a more accurate retrieval of the original feature vectors after encryption and decryption, preserving their fidelity and enhancing person re-identification accuracy. Secondly, this conversion to integers significantly reduces calculation error rates and encryption costs compared to traditional approaches. ECMO’s more efficient and precise process improves the scheme’s overall accuracy and practicality for real-world applications.

The tests assessed FREED’s efficiency compared to MGN, a well-used approach, in terms of computing and communication expenses. A considerable decrease in error rates was demonstrated for ECMO compared to other encoding techniques. Also established were the control parameter settings. FREED offered a secure and workable method for human re-identification, which outperformed earlier protocols in terms of computation and communication.

In conclusion, the article presents FREED, a novel and effective privacy-preserving person re-identification solution. By leveraging the Encoding Mechanism (ECMO) to convert floating-point feature vectors into integers, FREED addresses the limitations of traditional encoding methods, resulting in improved accuracy and reduced computation and calculation errors. The Secure Batch Multiplication (BatchSMUL) and Secure Batch Partial Decryption (BatchPDec) protocols enhance the system’s efficiency. Through extensive experimental evaluations, FREED demonstrates its effectiveness and efficiency compared to methods like MGN. Overall, FREED provides a promising approach to tackle the privacy challenges in person re-identification while maintaining high accuracy and practicality for real-world applications.