[Paper Review] Towards Evaluating Gaussian Blurring in Perceptual Hashing as a Facial Image Filter

Yigit Alparslan, Alparslan, Ken+2 more|arXiv (Cornell University)|Feb 1, 2020

Digital Media Forensic Detection10 references4 citations

TL;DR

This paper investigates whether applying Gaussian blurring before perceptual hashing improves detection of adversarial image manipulations—such as cropping, rotation, and text annotation—on facial images. Using the Yale B Extended Face Dataset, it finds that blurring only slightly improves accuracy (by 3.6% on cropped images), with DCT-based hashing outperforming average hashing for annotated images, but limited success across rotation attacks.

ABSTRACT

With the growth in social media, there is a huge amount of images of faces available on the internet. Often, people use other people's pictures on their own profile. Perceptual hashing is often used to detect whether two images are identical. Therefore, it can be used to detect whether people are misusing others' pictures. In perceptual hashing, a hash is calculated for a given image, and a new test image is mapped to one of the existing hashes if duplicate features are present. Therefore, it can be used as an image filter to flag banned image content or adversarial attacks --which are modifications that are made on purpose to deceive the filter-- even though the content might be changed to deceive the filters. For this reason, it is critical for perceptual hashing to be robust enough to take transformations such as resizing, cropping, and slight pixel modifications into account. In this paper, we would like to propose to experiment with effect of gaussian blurring in perceptual hashing for detecting misuse of personal images specifically for face images. We hypothesize that use of gaussian blurring on the image before calculating its hash will increase the accuracy of our filter that detects adversarial attacks which consist of image cropping, adding text annotation, and image rotation.

Motivation & Objective

• To evaluate the effectiveness of Gaussian blurring as a preprocessing step in perceptual hashing for detecting facial image misuse.
• To assess whether blurring improves robustness against common adversarial image manipulations like cropping, rotation, and text annotation.
• To compare the performance of different perceptual hashing techniques—DCT and average hashing—on facial images under adversarial conditions.
• To investigate the impact of image preprocessing (blurring) on hash similarity detection accuracy in a face-specific context.
• To determine if perceptual hashing can reliably flag unauthorized use of personal images despite minor visual perturbations.

Proposed method

• Utilized the Yale B Extended Face Dataset (576 poses per 28 subjects) for facial image experiments.
• Computed baseline perceptual hashes using Discrete Cosine Transform (DCT) and average hashing techniques on original images.
• Applied Gaussian blurring to images prior to hash computation to generate a second set of 'blurred' reference hashes.
• Generated adversarial test images via cropping (15% of original), 180° and 45° rotations, and text annotation (e.g., 'copyrighted').
• Computed perceptual hashes for all adversarial test images and compared them to both baseline and blurred reference hashes using Hamming distance.
• Evaluated detection accuracy across different hash bit-lengths (32-bit and 64-bit) and hash types (DCT vs. average) under varying threshold settings.

Experimental results

Research questions

• RQ1 Does applying Gaussian blurring before perceptual hashing improve the detection accuracy of adversarial facial image manipulations?
• RQ2 How do different perceptual hashing methods (DCT vs. average hashing) perform under various adversarial transformations (cropping, rotation, annotation)?
• RQ3 What is the impact of hash bit-length (32-bit vs. 64-bit) on detection accuracy for facial image duplicates?
• RQ4 To what extent does image rotation (180° and 45°) degrade perceptual hashing performance, and does blurring mitigate this?
• RQ5 Does text annotation significantly affect perceptual hash similarity, and can blurring improve resilience to such perturbations?

Key findings

• Gaussian blurring before hashing did not significantly improve overall detection accuracy across most adversarial transformations.
• The only measurable improvement was a 3.6% increase in accuracy for cropped images—rising from 14.3% (baseline) to 17.9% (blurred) with 64-bit average hashing.
• For annotated images, DCT-based hashing outperformed average hashing, achieving higher accuracy across threshold values.
• Detection accuracy for 180° rotated images reached 25–30% with 64-bit average hashing, but remained low for 45° rotations and DCT-based methods.
• For 45° rotated images, all hash types (32-bit and 64-bit DCT and average) achieved near 0% accuracy, indicating poor robustness.
• The 64-bit hash consistently outperformed 32-bit hashes in most cases, particularly for cropped and annotated images.