What the study found
The authors propose the first property-preserving hashing (PPH) construction for an ell1-distance predicate, which checks whether two one-sided ell1 distances between images stay within a threshold. They report that the scheme is highly efficient and has strong correctness guarantees.
Why the authors say this matters
The authors connect this work to detecting similar images under adversarial settings, where small imperceptible changes can make a perceptual hash fail. They suggest that choosing the threshold appropriately can force an attacker to add considerable noise to evade detection, which the abstract says would significantly deteriorate image quality.
What the researchers tested
The researchers designed a property-preserving hashing scheme for an ell1-distance predicate, extending prior PPH work that had focused on Hamming distance predicates. They evaluated runtime for grayscale images of size 28 x 28 and for larger RGB images of size 224 x 224, including a block-based setup for the larger images.
What worked and what didn't
The scheme is described as running in O(t^2) time. The abstract reports 0.0784 seconds for 28 x 28 grayscale images when pixel values were perturbed by up to 1%, and for 224 x 224 RGB images it reports 0.0128 seconds per block for 1% change and up to 0.2641 seconds per block for 14% change. The abstract does not report failures or cases where the scheme did not work.
What to keep in mind
The available summary does not describe experimental limitations beyond the stated image sizes, perturbation levels, and block-based evaluation for larger images. It also does not provide details on datasets, implementation conditions, or comparative baselines.
Key points
- The paper proposes the first property-preserving hashing construction for an ell1-distance predicate.
- The abstract says the scheme has strong correctness guarantees and is highly efficient.
- The authors link the work to detecting similar images under adversarial input attacks.
- Reported runtime is O(t^2), with 0.0784 seconds for 28 x 28 grayscale images at up to 1% perturbation.
- For 224 x 224 RGB images, the abstract reports 0.0128 seconds per block at 1% change and up to 0.2641 seconds per block at 14% change.
Disclosure
- Research title:
- New hashing scheme preserves ell1 distance predicates
- Authors:
- Hassan Jameel Asghar, Chenhan Zhang, Dali Kaafar
- Institutions:
- Macquarie University
- Publication date:
- 2026-04-21
- OpenAlex record:
- View
- Image credit:
- Photo by Tima Miroshnichenko on Pexels · Pexels License
Get the weekly research newsletter
Stay current with peer-reviewed research without reading academic papers — one filtered digest, every Friday.