Lightweight Blockchain-Assisted Image Authentication for Low-Bandwidth IoT Networks

Camera-equipped Internet of Things (IoT) nodes used for environmental monitoring, wildlife observation, industrial inspection and disaster response must deliver images whose integrity can be proven later, yet they transmit over links of a few tens of kilobits per second that lose frames in bursts. The common practice of hashing the complete image and anchoring that digest on a blockchain makes verification all-or-nothing: one lost frame invalidates the image, and ledger records grow with every camera and every frame. This paper proposes LBIA, a lightweight blockchain-assisted image authentication scheme that combines tile-aligned JPEG coding, a 64-bit truncated HMAC tag carried inside every tile, a signed Merkle root delivered as a sparse manifest, and Merkle batching of image roots so that one ledger transaction covers K images. A variant, LBIA-P, adds a 64-bit feature-level digest per tile for provenance after archival transcoding. LBIA was compared with full-image hashing, block hashing and feature-level authentication in a MATLAB-language simulation, executed in GNU Octave, on 12 test images of 512 x 512 pixels with IEEE 802.15.4-class framing and Bernoulli and Gilbert loss channels. Tile-aligned coding reproduced the whole-image reconstruction exactly and cost 0.39% extra rate for 64 x 64 tiles. In one-shot delivery at 10% bursty loss, LBIA authenticated 73.0% of the image area while full-image hashing authenticated none; with link-layer retransmission at 50 kbit/s and 5% bursty loss, the first verified tile arrived after 0.13 s instead of 9.65 s, for 3.94% more bytes on air. Batching 64 images per transaction cut ledger growth from 282 to 4.41 bytes per image at the cost of a 97.5 s anchoring delay. The feature-level digest reached an area under the ROC curve of 0.850 but missed every local-blur manipulation, so it is positioned as a complement to cryptographic tags rather than a replacement. Hardware validation, key management and adversarial collisions on the feature digest are left for future work.