Adaptive QR Streaming Framework

High-security environments — military systems, nuclear controls, industrial SCADA networks, and financial clearinghouses — use air-gapped networks: machines physically isolated from all RF and wired external links to prevent remote intrusion and data exfiltration. Even in an air-gapped system, administrators must still move data in (firmware updates, security patches, configuration files) and move data out (audit logs, diagnostics). Traditionally this relies on USB drives or optical media, which introduce the BadUSB class of hardware-level attacks and cumbersome media-management logistics.

Screen-Camera Communication (SCC) is a compelling alternative. Data is modulated into a sequence of visual frames on a transmitter screen and demodulated by a receiver camera. Because visible light cannot penetrate walls, the channel requires direct line-of-sight — eavesdropping is confined to a physically controlled space. And the appeal is broader than classified enclaves: a lecturer pushing a dataset to a room full of students, a field technician loading firmware onto an instrument with no radio, or two people exchanging a file without pairing, accounts, or an intermediary server.

Existing SCC research splits into two camps. Systems like ChromaCode (2018) and DeepLight (2021) pursue imperceptible communication — hiding data in subtle colour variations of ordinary video. While technically impressive, they require deep-learning decoders, are sensitive to ambient lighting, and offer lower reliability. The other camp — including this project — uses overt high-contrast QR codes, prioritising data integrity, robust decoding, and implementation simplicity. Yet even overt systems are rigid: fixed parameters that cannot adapt to different cameras, distances, or lighting. That adaptability gap is what this framework fills.

The channel is imperfect in ways that all scale with code density and frame rate. The rolling shutter in consumer cameras exposes a frame row by row, so a code that changes mid-exposure is captured torn — holding frames longer restores legibility. Perspective distortion compresses the apparent module size of codes near the edges of a dense grid first. Motion blur, autofocus hunting, and falling ambient contrast erode the smallest modules soonest. The practical capacity of the channel is therefore not a single number but a surface over the grid-size × frame-interval plane — the surface this study maps empirically.

Two principles anchor the design. The Nyquist–Shannon sampling theorem implies the receiver must oversample the display: to guarantee one clean capture of every frame held for duration t, the camera rate should comfortably exceed 2/t. And because the link is simplex (no acknowledgements), reliability comes from redundancy, not retransmission: the frame sequence loops cyclically and the receiver fills gaps on later passes — a lightweight form of fountain coding.

Present-day offline screen-camera systems are rigid and frequently inefficient: they give the operator no practical means of balancing transfer speed against delivery reliability, because neither the spatial layout (grid size) nor the temporal rate (frame interval) can be adjusted to the conditions at hand. The way grid layout (N×N) and frame interval (t) jointly govern the speed-versus-reliability curve has never been charted systematically.

QR codes (ISO/IEC 18004) encode data in a matrix of black and white modules with finder, alignment, and timing patterns that survive rotation and moderate perspective distortion. They give the framework four independent levers:

Reed-Solomon correction built into every QR reader repairs symbol-level damage for free; the framework layers a CRC-32 per chunk and a SHA-256 over the whole file on top, giving three independent guards against corruption while staying simple enough to run in a browser.

None of the surveyed systems exposes spatial and temporal parameters as first-class, user-tunable controls. A flagship-phone camera can resolve a dense 3×3 grid at close range; a modest laptop webcam may only decode a single code reliably. Exposing grid size and frame interval at run time lets one artifact span the whole hardware range — that hardware-aware tunability, rather than raw peak throughput, is the contribution.

The encrypted pages add an optional password layer on top of the standard protocol. On the sender: a random 16-byte salt and 12-byte IV are generated per transmission; the password is stretched through PBKDF2 (100 000 iterations, SHA-256) into a 256-bit AES-GCM key; the file is encrypted and the ciphertext is what gets chunked and streamed. The META frame carries the salt, IV, and a SHA-256 hash of the raw key bytes (the password verifier) — never the password.

On the receiver: the ciphertext is reassembled and SHA-256-verified first, proving transmission integrity independently of the password. The entered password is re-derived and checked against the verifier before any decryption is attempted, so a wrong password is caught instantly. AES-GCM's authentication tag then guarantees that a successful decryption is bit-perfect.

The study follows Design Science Research (Peffers et al., 2007): the artifact is built, demonstrated, and evaluated in iterative cycles. Evaluation is quantitative — the automated harness sweeps a full factorial matrix and logs telemetry to CSV for regression modelling.

The predictive model (RQ3) is validated, not merely fitted: runs are split into training and held-out test sets, accuracy reported as R², RMSE, and MAE, with k-fold cross-validation confirming the results don't depend on a particular split.

The framework is delivered as this zero-install web application, and the streaming core — segmentation, the self-describing protocol, encoding, detection, and reassembly — is being extracted into an open-source npm package (working name qr-stream) so any Node or JavaScript application can embed robust screen-camera transfer. Release accompanies the final research article.