Provenance Verification of AI-Generated Images via a Perceptual Hash Registry Using Merkle Patricia Trie and BK-Tree
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.
Nadia: We're looking at the paper: "Provenance Verification of AI-Generated Images via a Perceptual Hash Registry Using Merkle Patricia Trie and BK-Tree".
Elias: Here's the gist of the paper. The paper proposes a blockchain-backed framework for verifying AI-generated images through a registry-based provenance mechanism This system verifies whether an uploaded image matches a known, registered AI-generated image or its near-duplicate variants using perceptual hashing and similarity search anchored by blockchain commitments Core Mechanism The proposed system relies on a hybrid on-chain/off-chain architecture for secure, scalable verification This framework utilizes perceptual hashing (pHash) to create similarity-preserving fingerprints that are resilient to benign transformations like resizing and compression The hashes are stored using a Merkle Patricia Trie (MPT) for tamper-resistant storage on the blockchain and a Burkhard–Keller tree (BK) tree off-chain for efficient similarity search Verification Process The verification process involves several distinct stages when an image is uploaded to a digital platform The system first computes the pHash of the query image and uses it to extract a prefix key for lookup in the MPT structure Within the corresponding bucket, a BK-tree is queried based on Hamming distance to find potential matches If an exact match (Hamming distance = 0) or a match within a defined similarity threshold τ is found, the image is flagged as "Found in Registry" A normalized distance-based similarity score S(A, B) is also reported for near-duplicate matches where dH(A, B) = Hamming Distance and n = length of the hash Performance and Scalability The efficiency of the proposed indexing structure is demonstrated through performance analysis The system achieves stable and sublinear query times across registry sizes up to one million hashes, contrasting sharply with the linear growth seen in flat arrays or standalone BK-trees This performance gain stems from prefix-based partitioning, which bounds the candidate search space before metric-tree traversal begins Experimental Validation Experimental results confirm that the registry-based verification framework reliably identifies AI-derived images, including edited variants, while maintaining low false positive rates at conservative thresholds like τ = 6 Discontinuous prefixing is shown to be more robust against false positives compared to continuous prefixing under typical moderation constraints Furthermore, the results indicate that 2-bit prefix expansion is a more deployment-friendly default choice than 4-bit expansion due to its lower false positive rate Limitations and Future Directions A key limitation is the assumption that participating AI image generation platforms correctly register perceptual hashes at creation time The framework does not attempt to detect images from non-cooperative generators that do not participate in the registry Future work suggests integrating pHash with deep feature embeddings or exploring federated blockchain networks for enhanced multi-modal deepfake provenance The proposed system provides a transparent and tamper-evident registry mechanism that complements existing watermarking and learning-based detectors It establishes a scalable foundation for platform-level adoption of AI image provenance verification by jointly controlling robustness to benign transformations, false positive rates, and tail latency at scale The framework offers a practical, transparent, and deployable foundation for large-scale AI image provenance verification REFERENCES
1: Luo H, Li L, Li J.
Nadia: Digital Watermarking Technology for AI-Generated Images: A Survey.
Elias: Mathematics.
Nadia: 2025; 13(4):651
3: Zhao, X., Zhang, K., Su, Z., Vasan, S., Grishchenko, I., Kruegel, C., - & Li, L.
Paper summary: Nadia: So to recap, this paper proposes a blockchain framework for verifying AI images by assigning perceptual hashes at creation time, storing those on an on-chain Merkle Patricia Trie alongside an off-chain Burkhard–Keller tree for fast similarity searching.
Elias: And the thesis is that this registry mechanism gives you a tamper-proof way to check if an uploaded image matches registered AI content either exactly or within a set similarity threshold, reporting a similarity score along the way.
Priya: The system is designed so that AI platforms can compute their perceptual hash and record it on the blockchain right when they generate the image, which is key for establishing that initial provenance signal.
Nadia: And then when an image comes in later, you compute its hash, look it up in the Merkle Patricia Trie prefix structure to narrow down candidates, and then use a BK-tree search based on Hamming distance to find potential matches quickly.
Elias: The performance claims are pretty strong; they show stable and sublinear query times even when the registry grows up to one million hashes, which is much better than what you see in simpler structures.
Priya: But Elias, Priya wants to know about those numbers. What about the false positive rates when we use conservative thresholds, like a similarity threshold of six?
Nadia: The experiments confirm that this framework reliably identifies AI-derived images, including edited variants, while maintaining low false positive rates at that six threshold.
Elias: They also showed that discontinuous prefixing is more robust against false positives than continuous prefixing under typical moderation constraints, and they found that a two bit prefix expansion is the more deployment-friendly default choice over four bit expansion because it keeps the false positive rate lower.
Conclusion: Nadia: So looking at the authors, Apoorv Mohit Market Intelligence S andP Global Gurugram, Bhavya Aggarwal Market Intelligence S andP Global Gurugram, and Chinmay Gondhalekar Ratings S andP Global New York, United States of America. They are coming from a background in market intelligence and ratings.
Elias: It’s interesting how they approach image provenance from a very structured, data-centric viewpoint by using these specific tree structures to manage the massive scale of potential matches efficiently.
Priya: The implication for us is that this system offers a way to complement existing watermarking or learning-based detectors by providing a tamperproof signal for content authenticity in large online platforms.
Nadia: It moves the focus from just detecting an image as AI-generated to actually tracing its origin back through the registry, which adds a layer of verifiable proof to the process.
Elias: It establishes a scalable foundation for platform-level adoption because it handles both benign transformations and false positives in a way that keeps latency low at scale.
Priya: Ultimately, this work provides a practical, transparent mechanism for content provenance verification, even though we have to be mindful of the assumption that every AI image generator will cooperate by registering its content correctly.
S&P Global
cs.CR, cs.CY
Submitted: 2026-02-02
Updated: 2026-11-07
License: http://creativecommons.org/licenses/by-nc-sa/4.0/
Importance score: 90/100
The gist: The paper proposes a blockchain-backed framework for verifying AI-generated images through a registry-based provenance mechanism This system verifies whether an uploaded image matches a known…
Key concepts
- Perceptual Hashing (pHash)
- This technique creates a unique fingerprint for an image that remains consistent even if the image is slightly altered, such as being resized or compressed. It helps identify images that are visually similar, which is crucial for finding near-duplicates of AI-generated content.
- Merkle Patricia Trie (MPT)
- The MPT is a data structure used on the blockchain to store and organize hashes securely. It allows for tamper-proof storage of the registered image fingerprints, enabling efficient searching across millions of entries while maintaining security through cryptographic commitments.
- Burkhard–Keller Tree (BK-Tree)
- This is an off-chain search structure used to quickly find images similar to a query image. It organizes hashes based on their Hamming distance, allowing the system to efficiently locate potential matches within a defined similarity threshold.
- Registry-Based Provenance
- This framework establishes a centralized, verifiable list (registry) of known AI-generated images. When a new image is uploaded, the system checks if its hash matches an existing entry or is close enough to one, providing proof of provenance and authenticity.
Terminology
Summary
The paper proposes a blockchain-backed framework for verifying AI-generated images through a registry-based provenance mechanism This system verifies whether an uploaded image matches a known, registered AI-generated image or its near-duplicate variants using perceptual hashing and similarity search anchored by blockchain commitments
Core Mechanism
The proposed system relies on a hybrid on-chain/off-chain architecture for secure, scalable verification This framework utilizes perceptual hashing (pHash) to create similarity-preserving fingerprints that are resilient to benign transformations like resizing and compression The hashes are stored using a Merkle Patricia Trie (MPT) for tamper-resistant storage on the blockchain and a Burkhard–Keller tree (BK) tree off-chain for efficient similarity search
Verification Process
The verification process involves several distinct stages when an image is uploaded to a digital platform The system first computes the pHash of the query image and uses it to extract a prefix key for lookup in the MPT structure Within the corresponding bucket, a BK-tree is queried based on Hamming distance to find potential matches If an exact match (Hamming distance = 0) or a match within a defined similarity threshold τ is found, the image is flagged as Found in Registry
A normalized distance-based similarity score S(A, B) is also reported for near-duplicate matches where dH(A, B) = Hamming Distance and n = length of the hash
Performance and Scalability
The efficiency of the proposed indexing structure is demonstrated through performance analysis The system achieves stable and sublinear query times across registry sizes up to one million hashes, contrasting sharply with the linear growth seen in flat arrays or standalone BK-trees This performance gain stems from prefix-based partitioning, which bounds the candidate search space before metric-tree traversal begins
Experimental Validation
Experimental results confirm that the registry-based verification framework reliably identifies AI-derived images, including edited variants, while maintaining low false positive rates at conservative thresholds like τ = 6 Discontinuous prefixing is shown to be more robust against false positives compared to continuous prefixing under typical moderation constraints Furthermore, the results indicate that 2-bit prefix expansion is a more deployment-friendly default choice than 4-bit expansion due to its lower false positive rate
Limitations and Future Directions
A key limitation is the assumption that participating AI image generation platforms correctly register perceptual hashes at creation time The framework does not attempt to detect images from non-cooperative generators that do not participate in the registry Future work suggests integrating pHash with deep feature embeddings or exploring federated blockchain networks for enhanced multi-modal deepfake provenance
Conclusion
The proposed system provides a transparent and tamper-evident registry mechanism that complements existing watermarking and learning-based detectors It establishes a scalable foundation for platform-level adoption of AI image provenance verification by jointly controlling robustness to benign transformations, false positive rates, and tail latency at scale The framework offers a practical, transparent, and deployable foundation for large-scale AI image provenance verification
REFERENCES
[1] Luo H, Li L, Li J. Digital Watermarking Technology for AI-Generated Images: A Survey. p. 9 REFERENCES
[1] Luo H, Li L, Li J. Digital Watermarking Technology for
AI-Generated Images: A Survey. REFERENCES
[1] Luo H, Li L, Li J. Digital Watermarking Technology for
AI-Generated Images: A Survey. Digital Watermarking Technology for
AI-Generated Images: A Survey. Mathematics. Digital Watermarking Technology for
AI-Generated Images: A Survey. Mathematics.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[3] Zhao, X., Zhang, K. p. 6 We used the AI-Generated Images vs. Real Images dataset
[16], curated by Tristan Zhang and publicly available on
Kaggle. We used the AI-Generated Images vs. Real Images dataset
[16], curated by Tristan Zhang and publicly available on
Kaggle. , Zhang, K. , Vasan, S. , Zhang, K. , Vasan, S. , Vasan, S. , Grishchenko, I. , Vasan, S. , Grishchenko, I. , Grishchenko, I. , Kruegel, C. , Grishchenko, I. , Kruegel, C. , Kruegel, C. Invisible image watermarks are provably removable
using generative ai.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[4] de Ocáriz Borde, H. S. (2022). An overview of trees in blockchain technology: merkle trees and merkle patricia tries. p. 9 [5] de Ocáriz Borde, H. An overview of trees in blockchain technology: merkle trees and merkle patricia tries. [5] de Ocáriz Borde, H. An overview of trees in blockchain technology: merkle trees and merkle patricia tries.Where the paper says it
Where the paper says it
[5] Burkhard, W. A., & Keller, R. p. 0 Each AI-generated image is assigned a digital fingerprint
that preserves similarity using perceptual hashing and is registered at creation time by participating generation platforms. The hashes are stored on a hybrid on-chain/off-chain public
blockchain using a Merkle Patricia Trie for tamper-resistant
storage (on-chain) and a Burkhard–Keller tree (off-chain) to
enable efficient similarity search over large image registries. Each AI-generated image is assigned a digital fingerprint
that preserves similarity using perceptual hashing and is registered at creation time by participating generation platforms. The hashes are stored on a hybrid on-chain/off-chain public
blockchain using a Merkle Patricia Trie for tamper-resistant
storage (on-chain) and a Burkhard–Keller tree (off-chain) to
enable efficient similarity search over large image registries. , & Keller, R. Some approaches to bestmatch file searching. , & Keller, R. Some approaches to bestmatch file searching.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[6] Farid, H. (2021). An overview of perceptual hashing. p. 9 [7] Farid, H. An overview of perceptual hashing. [7] Farid, H. An overview of perceptual hashing.Where the paper says it
Where the paper says it
[8] Ahmed, N., Natarajan, T. p. 9 [8] Ahmed, N. , Natarajan, T. [8] Ahmed, N. , Natarajan, T. , Natarajan, T. , & Rao, K. , Natarajan, T. , & Rao, K. a) Perceptual Hashing: Perceptual hashing (pHash) generates a compact representation of an image by capturing its
dominant visual structures rather than exact pixel values. In
a typical pHash pipeline, the image is resized, converted to
grayscale, and transformed into the frequency domain using
a two-dimensional Discrete Cosine Transform (DCT) [8]. a) Perceptual Hashing: Perceptual hashing (pHash) generates a compact representation of an image by capturing its
dominant visual structures rather than exact pixel values. In
a typical pHash pipeline, the image is resized, converted to
grayscale, and transformed into the frequency domain using
a two-dimensional Discrete Cosine Transform (DCT) [8].Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[9] Heidari, A., Navimipour, N. p. 9 [9] Heidari, A. , Navimipour, N. [9] Heidari, A. , Navimipour, N. , Navimipour, N. , Talebi, S. , Navimipour, N. , Talebi, S. , Talebi, S. , & Unal, M. , Talebi, S. , & Unal, M. , & Unal, M. A novel blockchain-based deepfake detection method using federated
and deep learning models.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[10] Khalaf, L. I., Jumaili, M. p. 9 [10] Khalaf, L. , Jumaili, M. [10] Khalaf, L. , Jumaili, M. , Jumaili, M. , Almashhadany, M. , Jumaili, M. , Almashhadany, M. , Almashhadany, M. , Aljanabi, M. , Almashhadany, M. , Aljanabi, M. , Aljanabi, M. , Hasan, T. , Aljanabi, M. , Hasan, T. , Hasan, T. , & Algburi, S. , Hasan, T. , & Algburi, S.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[11] Mastoi, Q. U. A., Memon, M. p. 9 [11] Mastoi, Q. , Memon, M. [11] Mastoi, Q. , Memon, M. , Memon, M. , Jamil, A. , Memon, M. , Jamil, A. , Jamil, A. , Faique, M. , Jamil, A. , Faique, M.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[12] Farid, H. (2017). Reining in online abuses. p. 9 [12] Farid, H. Reining in online abuses. [12] Farid, H. Reining in online abuses.Where the paper says it
Where the paper says it
[13] Rosenthol, L. (2022, October). p. 9 [13] Rosenthol, L. (2022, October). [13] Rosenthol, L. (2022, October). (2022, October). C2PA: the world’s first industry standard
for content provenance (Conference Presentation). (2022, October). C2PA: the world’s first industry standard
for content provenance (Conference Presentation). C2PA: the world’s first industry standard
for content provenance (Conference Presentation). In Applications of
Digital Image Processing XLV (Vol. C2PA: the world’s first industry standard
for content provenance (Conference Presentation). In Applications of
Digital Image Processing XLV (Vol.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[14] Gowal, S., Bunel, R. p. 9 [14] Gowal, S. , Bunel, R. [14] Gowal, S. , Bunel, R. , Bunel, R. , Stimberg, F. , Bunel, R. , Stimberg, F. , Stimberg, F. , Stutz, D. , Stimberg, F. , Stutz, D. , Stutz, D. , Ortiz-Jimenez, G. , Stutz, D. , Ortiz-Jimenez, G. , Ortiz-Jimenez, G. , Kouridi,
C. , Ortiz-Jimenez, G. , Kouridi,
C. , Kouridi,
C. & Kohli, P. , Kouridi,
C. & Kohli, P. & Kohli, P. SynthID-Image: Image watermarking at
internet scale. SynthID-Image: Image watermarking at
internet scale. arXiv preprint arXiv:2510.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[15] Lashkari, B., & Musilek, P. p. 9 [15] Lashkari, B. , & Musilek, P. [15] Lashkari, B. , & Musilek, P. , & Musilek, P. A comprehensive review of
blockchain consensus mechanisms. , & Musilek, P. A comprehensive review of
blockchain consensus mechanisms.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[16] https://www.kaggle.com/datasets/tristanzhang32/ai-generated-imagesvs-real-images
[17] Gent, E. (2023). A Cryptocurrency for the Masses or a Universal ID? p. 9 com/datasets/tristanzhang32/ai-generated-imagesvs-real-images
[17] Gent, E. A Cryptocurrency for the Masses or a Universal ID? com/datasets/tristanzhang32/ai-generated-imagesvs-real-images
[17] Gent, E. A Cryptocurrency for the Masses or a Universal ID? A Cryptocurrency for the Masses or a Universal ID? :
Worldcoin Aims to Scan all the World’s Eyeballs. A Cryptocurrency for the Masses or a Universal ID? :
Worldcoin Aims to Scan all the World’s Eyeballs.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[18] Park, N., & Kim, S. p. 9 [18] Park, N. , & Kim, S. [18] Park, N. , & Kim, S. , & Kim, S. How do vision transformers work? , & Kim, S. How do vision transformers work? How do vision transformers work? arXiv
preprint arXiv:2202. How do vision transformers work? arXiv
preprint arXiv:2202.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[19] Nguyen, C. T., Hoang, D. p. 9 [19] Nguyen, C. , Hoang, D. [19] Nguyen, C. , Hoang, D. arXiv
preprint arXiv:2202. [19] Nguyen, C. arXiv
preprint arXiv:2202. [19] Nguyen, C. , Nguyen, D. ,
Dutkiewicz, E. , Nguyen, D. ,
Dutkiewicz, E. ,
Dutkiewicz, E. , & Tuong, N. ,
Dutkiewicz, E. , & Tuong, N.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[20] Luo H, Li L, Li J. Digital Watermarking Technology for AI-Generated Images: A Survey. Mathematics. 2025; 13(4):651
[21] Lokner Ladevi ¯ c, A., Kramberger, T. p. 9 3390/math13040651
[2] Lokner Ladevi ¯ c, A. , Kramberger, T. 3390/math13040651
[2] Lokner Ladevi ¯ c, A. , Kramberger, T. , Kramberger, R. , & Vlahek, D. , & Vlahek, D. Detection of AI-generated synthetic images with a lightweight
CNN.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[22] Zhao, X., Zhang, K., Su, Z., Vasan, S., Grishchenko, I., Kruegel, C., - & Li, L. (2024). Is Artificial Intelligence Generated Image Detection a Solved Problem? p. 9 , Xiong, L. Is Artificial Intelligence Generated Image Detection a Solved Problem? Is Artificial Intelligence Generated Image Detection a Solved Problem? arXiv preprint arXiv:2505.Where the paper says it
Where the paper says it
[23] Farid, H. (2017). Reining in online abuses. Technology and Innovation, 19(3), 249-255
[24] Rosenthol, L. (2022, October). C2PA: the world’s first industry standard for content provenance (Conference Presentation). In Applications of Digital Image Processing XLV (Vol. 12226, p. 122260P)
[5] de Ocáriz Borde, H. S. (2022). An overview of trees in blockchain technology: merkle trees and merkle patricia tries. University of Cambridge: Cambridge, UK
[6] Burkhard, W. A., & Keller, R. M. (1973). Some approaches to bestmatch file searching. Communications of the ACM, 16(4), 230-236
[7] Farid, H. (2021). An overview of perceptual hashing. Journal of Online Trust and Safety, 1(1)
[8] Ahmed, N., Natarajan, T., & Rao, K. R. (2006). Discrete cosine transform. IEEE transactions on Computers, 100(1), 90-93
[9] Heidari, A., Navimipour, N. J., Dag, H., Talebi, S., & Unal, M. (2024). A novel blockchain-based deepfake detection method using federated and deep learning models
[10] Khalaf, L. I., Jumaili, M. L. F., Almashhadany, M. T. M., Aljanabi, M. S., Hasan, T. S., & Algburi, S. (2024). Image forgery detection using convolutional neural networks and blockchain technology
[11] Mastoi, Q. U. A., Memon, M. F., Jan, S., Jamil, A., Faique, M., Ali, Z., - & Syed, T. A. (2025). Enhancing Deepfake Content Detection Through Blockchain Technology
[13] Rosenthol, L. (2022). C2PA: the world’s first industry standard for content provenance (Conference Presentation)
[14] Gowal, S., Bunel, R., Stimberg, F., Stutz, D., Ortiz-Jimenez, G., Kouridi, C., - & Kohli, P. (2025). SynthID-Image: Image watermarking at internet scale
[15] Lashkari, B., & Musilek, P. (2021). A comprehensive review of blockchain consensus mechanisms. IEEE access, 9, 43620-43652
[16] https://www.kaggle.com/datasets/tristanzhang32/ai-generated-imagesvs-real-images
[17] Gent, E. (2023). A Cryptocurrency for the Masses or a Universal ID?: Worldcoin Aims to Scan all the World’s Eyeballs. IEEE Spectrum, 60(1), 42-57
[18] Park, N., & Kim, S. (2022). How do vision transformers work?. arXiv preprint arXiv:2202.06709
[19] Nguyen, C. T., Hoang, D. T., Nguyen, D. N., Xiao, Y., Pham, H. A., Dutkiewicz, E., & Tuong, N. H. (2023).
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved AI systems can integrate provenance mechanisms directly into their pipelines by computing a perceptual hash of each generated image and recording it on a blockchain registry at creation time. This allows platforms to
register content at creation time by computing the perceptual hash of each generated image and recording it on the blockchain registry.
-
The system can perform verification when images are re-uploaded to digital platforms, enabling
identification of previously registered AI-generated images even after benign transformations or partial modifications.
This provides a deterministic and auditable mechanism for identifying re-uploads. -
The framework allows for similarity verification using perceptual hashing and Hamming distance, which is
resilient to benign transformations such as resizing, compression, or minor brightness changes.
This enables reliable identification even when images undergo slight modifications. -
The system can utilize a Merkle Patricia Trie (MPT) organized by prefix to enable
efficient insertion and retrieval by narrowing the search space to a manageable subset of visually similar images.
This structure supports scalable storage and efficient indexing. -
The system employs a Burkhard–Keller tree within each MPT bucket to support
fast identification of perceptually similar images based on Hamming distance.
This allows for approximate nearest-neighbor search, enabling the identification of near-duplicate variants at upload time. -
The verification process can be tuned using an operational Hamming distance threshold tau to balance true positive and false positive rates, as described by the finding that
an operating point around tau = 6 consistently recovers approximately 98–99% of AI-derived content with FPR below 0.2%.
-
The architecture minimizes on-chain storage to cryptographic commitments only, as the system stores
full perceptual hashes and their associated metadata such as image creation timestamp, generating platform identifier, and optional image signatures off-chain within BK-tree structures.
This design ensures scalability while maintaining tamper-evidence through on-chain root hashes. -
The system can provide a normalized distance based similarity score S(A, B) to quantify the degree of similarity between an uploaded image and a registered one, where
S(A,B) = 1 − dH(A, B) / n × 100.
This score is not intended to represent calibrated probability but serves as a heuristic for flagging near-duplicate content.
Abstract
The rapid advancement of artificial intelligence has made the generation of synthetic images widely accessible, increasing concerns related to misinformation, digital forgery, and content authenticity on large-scale online platforms. This paper proposes a registry-based framework for verifying AI-generated images through a registry-based provenance mechanism. Each AI-generated image is assigned a digital fingerprint that preserves similarity using perceptual hashing and is registered at creation time by participating generation platforms. The Merkle Patricia Trie, provides prefix-based organization and cryptographic integrity of the registry structure, while Burkhard-Keller tree enable similarity search based on Hamming distance. Verification is performed when images are re-uploaded to digital platforms such as social media services, enabling identification of previously registered AI-generated images even after benign transformations or partial modifications. The proposed system does not aim to universally detect all synthetic images, but instead focuses on verifying the provenance of AI-generated content that has been registered at creation time. By design, this approach complements existing watermarking and learning-based detection methods, providing a platform-agnostic, tamper-evident mechanism for scalable content provenance and verification at the point of large scale online distribution.
Sources
- Is Artificial Intelligence Generated Image Detection a Solved Problem?
- SynthID-Image: Image watermarking at internet scale
- How Do Vision Transformers Work?
Related papers
- From Strategy Search to Policy Learning: MCTS-Guided Adversarial Co-Evolution for LLM Safety Alignment
- Privacy Leakage from Recurrent State in LLM Inference
- Incrementally Verifiable Oracle Aided Computation
- Privilege and confidentiality in generative AI workflows
- BadMoE: Backdooring Mixture-of-Experts LLMs via Routing-Aware Triggers and Task-Underutilized Experts
- CoVault: Secure, Scalable Analytics of Personal Data