A multi-level privacy protection method based on bimodal PSNR quality control

Abstract

A multi-level privacy protection method based on bimodal PSNR quality control. A bimodal PSNR evaluation system is constructed to separate the calculation of Face-PSNR and Background-PSNR. An adaptive blur kernel is used for dynamic desensitization, and the kernel is calculated as max(w, h) / 6 based on the ROI size. A PSNR-driven iterative blur mechanism is introduced, and each round is checked until it meets the standard and is backed up with a mosaic. Differential privacy protection is applied to the log, and a heterogeneous detector is used to verify the desensitization effect independently. An accuracy attenuation check and compliance audit mechanism is established. This method achieves real-time processing at 17.8 FPS on Raspberry Pi 4B, with a desensitization incremental delay of 3.6 ms, ROIPSNR<22 dB and global PSNR>35 dB, accuracy attenuation≤3.41%, and privacy unidentifiability rate of 100%, meeting the GDPR / PIPL compliance requirements. It is suitable for privacy protection target detection in community security edge computing scenarios.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, privacy protection, and edge computing, and more specifically, to a multi-level privacy protection method for community security scenarios. Background Technology

[0002] With the deepening of smart city construction, video surveillance systems are playing an increasingly important role in the field of public safety. However, traditional security monitoring systems face serious privacy protection compliance challenges: 1. Inadequate privacy protection mechanisms: Traditional Gaussian blur or mosaic uses fixed parameters and cannot adapt to the dynamic changes in pedestrian scale in videos. Experimental data shows that when the global PSNR=25dB, faces can still be recognized by the ArcFace algorithm with an accuracy of more than 85%, while the background is excessively blurred, causing the pedestrian detection mAP50 to drop by more than 15%.

[0003] 2. Lack of quantitative quality assessment of desensitization: Most existing solutions rely on subjective judgment that the desensitization effect "looks blurry" without objective indicators to verify the desensitization effect, and cannot provide regulatory agencies with an auditable chain of compliance evidence.

[0004] 3. The conflict between privacy protection and detection accuracy: Excessive desensitization can impair detection capabilities, while insufficient desensitization can lead to privacy leaks. Existing solutions lack a quantitative assessment mechanism for accuracy degradation.

[0005] 4. Lack of edge deployment verification: More than 90% of existing research focuses on server-side simulation and has not verified the real-time overhead of de-identification operations on real edge devices such as Raspberry Pi.

[0006] 5. Security risks of log data: The de-identified logs contain sensitive information such as detection box coordinates and PSNR values. Attackers can analyze the logs to trace the attacker's location, creating a secondary privacy leakage channel. Summary of the Invention

[0007] S1. Construct a bimodal PSNR privacy assessment system: Separate the region of interest (ROI) and the background region from the input image, calculate Face-PSNR and Background-PSNR respectively, and establish the privacy strength objective function Privacy_Score = α×(1 / PSNR_roi) + β×(1 / PSNR_global), where α=0.6 and β=0.4. The optimal parameter combination satisfies Face-PSNR≈20dB and Background-PSNR∈[35,45]dB; S2, Adaptive blur kernel dynamic desensitization: Gaussian blur and mosaic dual desensitization are adopted. The blur kernel size is adaptively calculated based on the ROI size, kernel=max(w,h) / 6, and the blur area of ​​occluded and side face samples is expanded by 15%. S3, PSNR-driven iterative blurring mechanism: Gaussian blurring is applied to the ROI region round by round and PSNR is calculated in real time. If PSNR_roi>22dB, the blurring rounds are added, with a maximum of 10 iterations. Finally, a 7×7 mosaic block is used as a backup guarantee. S4. Differential Privacy Log Protection: Gaussian noise is applied to sensitive numerical fields in the de-identified logs, and a (ε,δ)-differential privacy mechanism is adopted, with a total privacy budget of ε=1.0. S5. Independent verification by heterogeneous detectors: Using the OpenCV Haar cascade detector as an independent verifier, the facial region of the desensitized image is rescanned to ensure a privacy non-identification rate of ≥95%; S6. Accuracy Attenuation Verification and Compliance Audit: Compare the pedestrian detection accuracy of the original image and the desensitized image, calculate the mAP50 attenuation rate, establish a GDPR / PIPL compliance checklist, and generate an auditable chain of evidence.

[0008] The beneficial effects of this invention are: 1. The dual-peak PSNR mechanism fills the gap in the quantification of privacy intensity: It simultaneously constrains the "lower limit of background sharpness (>35dB)" and the "upper limit of facial blur (<22dB)" on the same image, and improves the quality of privacy desensitization from subjective judgment to a quantifiable and auditable engineering indicator.

[0009] 2. Adaptive fuzzy kernel solves the multi-scale adaptation problem: Based on the ROI size, the fuzzy kernel is dynamically allocated, the maximum ROIPSNR value is reduced from 26.27dB to 21.38dB, the global PSNR average is improved by 1.8dB, and a 100% compliance rate is achieved.

[0010] 3. PSNR-driven iterative fuzzing ensures zero violations: an average of only 1.3 iterations are required, and the desensitization incremental delay is only 3.6ms.

[0011] 4. Differential privacy protection blocks log leakage channels: 130 records consume only 4.23% of the privacy budget, and attackers cannot reverse engineer the original data through the logs.

[0012] 5. Heterogeneous validation avoids self-validation bias: 100% pass rate for 130 images tested.

[0013] 6. Three-dimensional joint optimization: Real-time processing at 17.8 FPS on Raspberry Pi 4B with accuracy attenuation ≤3.41% and ROIPSNR <22dB for all, meeting the "de-identification" requirement of Article 51 of the Personal Information Protection Law.

[0014] 7. Auditable chain of evidence meets compliance audit requirements: All 10 items on the GDPR / PIPL compliance checklist have been passed. Attached Figure Description

[0015] In the picture: 1. Input Image Module 2. ROI Extraction Unit 3. Background Separation Unit 4. Face-PSNR Calculation Unit 5. Background-PSNR Calculation Unit 6. Privacy Score Calculation Unit 7. Adaptive Blur Kernel Calculation Unit 8. Gaussian Blur Processing Unit 9. Mosaic Processing Unit 10. PSNR Iterative Checking Unit 11. Differential Privacy Noise Addition Unit 12. Log Storage Unit 13. Haar Cascade Detector 14. Independent Verification Unit 15. Accuracy Comparison Unit 16. Compliance Audit Unit 17. Archive Package Generation Unit The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0016]

[0017] Figure 1 Figure 1 The accompanying figure is a summary of the present invention, illustrating the principle of the bimodal PSNR evaluation system of the present invention. The input image (1) is fed into the ROI extraction unit (2) and the background separation unit (3) in parallel; the ROI extraction unit (2) uses a Haar detector to locate the face and expands the bounding box by 1.5 times, and obtains the face-PSNR through the Face-PSNR calculation unit (4); the background separation unit (3) obtains the background-PSNR through the Background-PSNR calculation unit (5); finally, the Privacy_Score is calculated by the Privacy Score calculation unit (6).

[0018]

[0019] Figure 2 Figure 2 This is the flowchart of the adaptive fuzzy kernel dynamic desensitization process of the present invention. The adaptive fuzzy kernel calculation unit (7) calculates kernel=max(w,h) / 6 according to the ROI size, and then the Gaussian fuzzing processing unit (8) applies a 61×61 kernel (σ=18) for processing. Then the mosaic processing unit (9) superimposes a 7×7 block size for processing. Finally, the PSNR iteration checking unit (10) verifies that the standard is met.

[0020]

[0021] Figure 3 Figure 3This is the five-module pipeline flowchart of the present invention. Dynamic desensitization processing (S1-S3) outputs to differential privacy log protection (S4), sensitive information verification (S5), and precision attenuation verification (the first three items of S6); the outputs of each step are summarized to compliance audit archiving (the last item of S6) to generate an archive package.

[0022]

[0023] Figure 4 Figure 4 This is a diagram of the PSNR-driven iterative fuzzing mechanism of the present invention. After the first round of fuzzing, the PSNR is checked. If it is >22dB, additional fuzzing rounds are added, up to a maximum of 10 rounds, and finally, a mosaic effect is used as a final buffer.

[0024]

[0025] Figure 5 Figure 5 This is a schematic diagram of the differential privacy protection principle of the present invention. The original log values ​​are stored in the log storage unit (12) after Gaussian noise is applied by the differential privacy noise addition unit (11).

[0026]

[0027] Figure 6 Figure 6 This is a diagram of the heterogeneous detector independent verification mechanism of the present invention. The desensitized image is input in parallel to the YOLO detector and the Haar cascade detector (13), and the latter outputs an objective verification conclusion through the independent verification unit (14). Detailed Implementation

[0028] Please see Figures 1 to 6 The embodiments of the present invention include: Example 1: Construction of a Bimodal PSNR Privacy Assessment System The input image (1) enters the ROI extraction unit (2), where the face is located using an OpenCV Haar cascade detector and the ROI is enlarged by 1.5 times. The Face-PSNR calculation unit (4) and the Background-PSNR calculation unit (5) calculate the PSNR of the two regions respectively by 10log. 10 (MAX² / MSE). The privacy score calculation unit (6) is calculated as Privacy_Score = 0.6×(1 / PSNR_roi)+ 0.4×(1 / PSNR_global). Validation of 130 CityPersons subset images: global PSNR mean 37.83dB, ROIPSNR mean 17.04dB, compliance rate 99.0%.

[0029] Example 2: Adaptive Fuzzy Kernel Dynamic Desensitization The adaptive blur kernel calculation unit (7) is allocated according to the ROI size: kernel=15 when ROI<90px, 15-25 when 90-150px, 25-45 when 150-270px, and 45-61 when >270px. The area of ​​occlusion and side face is expanded by 15%. The Gaussian blur processing unit (8) applies a 61×61 kernel (σ=18) and superimposes it three times. The mosaic processing unit (9) superimposes 7×7 blocks. ShanghaiTech video 500 frames verification: the maximum value of ROIPSNR drops to 21.38dB, 100% meets the standard, and the global PSNR average is 38.45dB.

[0030] Example 3: PSNR-driven iterative fuzzy mechanism The PSNR iterative checking unit (10) checks each round. If PSNR_roi>22dB, blurring is added (sigma +30% per round), up to a maximum of 10 rounds. If the standard is not met, mosaic is used as a fallback. On average, it takes 1.3 rounds, and 78% of single rounds meet the standard. The desensitization incremental delay is 2.1ms (PC) / 3.6ms (Raspberry Pi).

[0031] Example 4: Differential Privacy Log Protection The differential privacy noise addition unit (11) applies Gaussian noise to the PSNR value, processing time, and detection box coordinates, with σ_noise=Δ / ε. The total budget ε=1.0, 130 records consume 4.23%, and the coordinate offset is <2 pixels.

[0032] Example 5: Independent Validation of Heterogeneous Detectors The Haar cascade detector (13) uses Haar features + AdaBoost, scaleFactor=1.2, minNeighbors=6, minSize=40×40. The independent validation unit (14) calculates the Laplacian variance score, and blur_score>50 is marked as unqualified. The pass rate of 130 images is 100%.

[0033] Example 6: Precision Attenuation Verification and Compliance Audit The accuracy comparison unit (15) compared the original and desensitized detection results using IoU matching (threshold = 0.5). The mAP50 attenuation rate was 3.23% / 2.87% / 3.41% (three stages), and the key point error was 1.48px / 1.35px / 1.46px, all of which met the standards. The compliance audit unit (16) checked each item against Articles 5 / 17 / 25 / 32 of the GDPR and Articles 6 / 25 / 47 / 51 of the PIPL, and all 10 items passed. The archive package generation unit (17) output privacy_protection_archive.zip (2.36MB).

[0034] Example 7: Raspberry Pi 4B End-Side Deployment Verification This method was deployed on a Raspberry Pi 4B (BCM2711, 4×Cortex-A72@1.5GHz, 4GB RAM), using the NCNNFP32 framework with 640×480 input. NCNN inference took 49.1ms, preprocessing 3.2ms, face estimation 0.3ms, Gaussian blur 2.8ms, mosaic 0.5ms, and post-processing 0.3ms, for a total of 56.2ms (17.8 FPS). Desensitization increment was 3.6ms (6.4%), memory usage was 198MB, and power consumption was 2.1W. In a 100-frame continuous stream: the global PSNR average was 38.12dB (std=2.41), the ROIPSNR average was 16.89dB (std=2.73), and the maximum was 21.85dB, achieving 100% compliance.

Claims

1. A multi-level privacy protection method based on bimodal PSNR quality control, characterized in that, Includes the following steps: S1. Construct a bimodal PSNR privacy assessment system: Separate the region of interest (ROI) and the background region from the input image, calculate Face-PSNR and Background-PSNR respectively, and establish the privacy strength objective function Privacy_Score = α×(1 / PSNR_roi) + β×(1 / PSNR_global), where α=0.6 and β=0.

4. The optimal parameter combination satisfies Face-PSNR≈20dB and Background-PSNR∈[35,45]dB; S2, Adaptive blur kernel dynamic desensitization: Gaussian blur and mosaic dual desensitization are adopted. The blur kernel size is adaptively calculated based on the ROI size, kernel=max(w,h) / 6, and the blur area of ​​occluded and side face samples is expanded by 15%. S3, PSNR-driven iterative blurring mechanism: Gaussian blurring is applied to the ROI region round by round and PSNR is calculated in real time. If PSNR_roi>22dB, the blurring rounds are added, with a maximum of 10 iterations. Finally, a 7×7 mosaic block is used as a safety net to ensure that ROIPSNR<22dB and global PSNR>35dB. S4. Differential Privacy Log Protection: Gaussian noise is applied to sensitive numerical fields in the de-identified logs, and a (ε,δ)-differential privacy mechanism is adopted, with a total privacy budget of ε=1.

0. S5. Independent verification by heterogeneous detectors: Using the OpenCV Haar cascade detector as an independent verifier, the facial region of the desensitized image is rescanned to ensure a privacy non-identification rate of ≥95%; S6. Accuracy Attenuation Verification and Compliance Audit: Compare the pedestrian detection accuracy of the original image and the desensitized image, calculate the mAP50 attenuation rate, establish a GDPR / PIPL compliance checklist, and generate an auditable chain of evidence.

2. The method according to claim 1, characterized in that, The bimodal PSNR evaluation system in S1 uses an OpenCV Haar cascade detector to locate the face ROI. The region of interest is defined by expanding the face bounding box by 1.5 times. The background region is the image area outside the ROI. The peak signal-to-noise ratio (PSNR) of the two regions is calculated as 10log[…]. 10 (MAX² / MSE).

3. The method according to claim 1, characterized in that, In S2, the adaptive blur kernel is dynamically allocated according to the ROI size: when the ROI size is <90px, kernel=15; when it is 90-150px, kernel=15-25; when it is 150-270px, kernel=25-45; and when it is >270px, kernel=45-61, ensuring that the blur energy is concentrated inside the ROI.

4. The method according to claim 1, characterized in that, In the S3, the average number of iterations for PSNR-driven iterative blurring is 1.3, and 78% of the ROI is achieved in a single iteration. The desensitization incremental delay is controlled within 3.6ms / frame.

5. The method according to claim 1, characterized in that, In the S4 differential privacy protection, the standard deviation of Gaussian noise σ_noise=Δ / ε, where Δ is the field sensitivity and ε is the privacy budget allocation value. The actual privacy budget consumption for 130 records is only 4.23%.

6. The method according to claim 1, characterized in that, The heterogeneous detector independent verification in S5 adopts a different algorithm principle than the desensitization engine. The Haar feature + AdaBoost classifier and the YOLO detector have no shared bias, thus avoiding self-verification bias.

7. The method according to claim 1, characterized in that, The precision attenuation verification in S6 compares the original detection results with the desensitized detection results through IoU matching. The compliance standards are mAP50 attenuation rate ≤5%, joint point error <3px, and all 10 items of the GDPR / PIPL compliance checklist are passed.

8. The method according to claim 1, characterized in that, The method is deployed on a Raspberry Pi 4B edge device, achieving real-time processing of 17.8 FPS under the NCNN FP32 inference framework, with a desensitization incremental latency of 3.6 ms, memory usage of ≤200 MB, and power consumption of ≤3 W.

Images