A camera owner today can share encrypted footage in two ways. They can share it by time ("Tuesday 9 to 5"), which is cheap and provably safe. Or they can share it by content ("only the parts showing the delivery driver"), which requires trusting a server that looks at the plaintext. We propose a third option. Each video segment is tagged on the owner's device, at recording time, with concepts from a learned hierarchy. Its encryption key is derived from both a time tree and a concept tree. A recipient handed a few keys can then decrypt exactly "vehicles at the driveway, last week" and nothing else, with no server ever seeing the video. Owners write sharing policies in their own language, including abstract ones such as "anything private" or "children playing." A multilingual vision-language encoder compiles each policy into a small set of nodes in the concept tree.
The cryptography in this design is simple. The hard part is in machine learning. Once concepts are fixed into keys at encryption time, every tagging mistake becomes an access-control mistake: a spurious tag leaks a segment to someone who should not see it, and a missed tag denies a segment to someone who should. The shape of the concept hierarchy decides how often those mistakes happen. Our central claim is that the hierarchy should be learned jointly from visual confusability and cross-lingual policy phrasing, rather than taken from a fixed lexical taxonomy. A learned hierarchy should give lower leak rates, at equal denial rates and equal key cost, than fixed taxonomies or translate-to-English pipelines. The advantage should be largest for abstract concepts and low-resource languages.
Where this sits
Hierarchical key derivation means deriving many keys from one root by repeatedly applying a one-way hash. Anyone holding a node's key can compute all keys below it but none above or beside it. This line of work supplies our mechanism. CaCTUs builds a binary tree whose leaves are 10-second epochs of encrypted camera video, so that one internal node delegates a time window; it runs on a Raspberry Pi 4 at 480p and 10 fps with about 2 seconds of latency. Backes et al. formalize such trees for lazy revocation, and Huang et al. report that on-the-fly derivation cuts key storage to roughly 20% of storing every key independently. All three partition a single linear axis. We keep the time tree from CaCTUs unchanged and add a second, semantic axis.
The closest prior system is MD-SMAC. It applies attribute-based encryption, in which a decryption key works only if its holder's attributes satisfy a policy, across three nested dimensions of scalable video: signal quality, resolution and frame rate. Those dimensions are nested by the codec bitstream, so containment is exact by construction. Our concept dimension has no such guarantee. Nesting is a modeling decision that a classifier must uphold, and that is the gap this proposal addresses.
GRAM gates concepts by dropping neural modules at inference time. It recovers to about 56% of baseline knowledge under adversarial fine-tuning, versus about 87% for MaxEnt unlearning, which shows residual leakage. It also protects model capabilities, not stored data. AgentGuard interprets abstract intents with an LLM but enforces policy only procedurally on a server. Compass does encrypted semantic search, but for a single user and with about 500 MB of client hints. None of these lets a data owner delegate content-defined subsets of encrypted media to multiple people offline.
For grounding, we reuse multilingual encoders as the tagging and policy-compilation engine. V-SONAR projects a frozen vision encoder into a text space covering more than 1,500 languages. AltCLIP swaps CLIP's text encoder for a multilingual one. MURAL anchors under-resourced languages through translation pairs.
Three findings motivate the learned hierarchy, and they also constrain it. MURAL notes that the "images are a universal pivot" assumption fails on non-visual concepts. V-SONAR's visual alignment is capped by the quality of its text space in low-resource languages. Iosif et al. show that abstract nouns have diffuse, attribute-based neighborhoods rather than tight clusters. Together these predict that an abstract policy will not map onto a single tree node, and that its mapping may drift across languages.
Core Claim
The concept tree is the unit of risk. Each segment is wrapped under every concept leaf whose tag probability exceeds a threshold. A delegate holding a set of nodes $C$ decrypts every segment wrapped under any leaf below $C$. For a distribution of multilingual policies $\pi$, each compiled to a node cover $C_\pi$, we choose the hierarchy $\mathcal{T}$ and the tagger thresholds to minimize the following objective:
$$ \mathbb{E}_{\pi}\big[\,\mathrm{Leak}(\pi;\mathcal{T}) + \lambda\,\mathrm{Deny}(\pi;\mathcal{T}) + \mu\,|C_\pi|\,\big] $$
The terms are defined as follows. Leak is the fraction of released segments that do not satisfy $\pi$ under human ground truth. Deny is the fraction of satisfying segments that are not released. $|C_\pi|$ is the number of node keys needed for one time window, which sets delegation cost. The weights $\lambda$ and $\mu$ trade denial and key cost against leakage.
The hypothesis follows from the objective. Concepts that are visually confusable but semantically far apart should sit far apart in the tree only if they really are far apart in policy space. Concepts that policies in many languages treat as a unit should share a subtree. A fixed taxonomy optimizes neither property.
Research Program
The first component is the derivation scheme. Concept keys are derived top-down by hashing the parent key with the child's label. Each segment's content key is derived in a fixed order: first down the concept tree, then down that concept's time tree. A delegate for concept node $v$ over window $\tau$ therefore receives one time-cover key per concept leaf under $v$. That is $|\mathrm{leaves}(v)|\cdot O(\log N)$ keys, where $N$ is the number of epochs. Fixing the order matters because it prevents two delegates from pooling keys to unlock combinations of concept and time that neither was given. The price is that delegation cost grows with subtree width, so tree shape directly enters the $|C_\pi|$ term.
The second component is the tagger. A frozen multilingual video encoder (V-SONAR as the main choice, AltCLIP for images) runs on the device. A lightweight head on top of it is calibrated per leaf. We include relation-aware pooling in the spirit of RDAN, which aligns verbs and adjectives across clusters of co-occurring regions, because action-like and abstract concepts are not confined to one box.
The third component learns the hierarchy. We start from a large concept vocabulary and build $\mathcal{T}$ by agglomerative merging. The merge score combines three quantities: the tagger's confusion between two concepts on held-out data; how often the two concepts appear in the same policy, measured from policy phrases written by native speakers in many languages; and a cross-lingual consistency term that penalizes merges which make one policy compile to different covers in different languages. We then fine-tune the thresholds under the objective above.
The fourth component is policy compilation. A policy phrase in any language is embedded with the same encoder and compiled to a minimal node cover by greedy set cover over node centroids. The owner then confirms the cover against thumbnails of their own footage, which stay on the device. The confirmation step lets the owner catch a bad compilation before any keys leave the device.
Proposed Experiments
No benchmark exists for content-delegated encrypted media; we found no reported result on Lacuna for concept-level delegation leak or denial rates. We will therefore construct one from existing multilingual captioned data. Video comes from VATEX (English and Chinese captions), DREAM-1K and PE-VIDEO. Images come from Multi30K (English, German, French and Czech). Policy phrasing in more languages draws on M3IT.
Ground truth for whether a segment satisfies a concept will come from its human captions. An LLM proposes judgments, and humans audit a stratified sample. The policy set will span concrete concepts ("bicycle"), actions ("cooking") and abstract or social concepts ("intimacy," "danger," "celebration"). Concreteness will be stratified by the neighborhood-diffuseness diagnostic of Iosif et al. computed in the encoder's text space.
Metrics are leak rate at fixed denial (5% and 10%), the full leak–denial curve, keys per delegation, cross-lingual cover agreement (the Jaccard overlap of the covers compiled from translations of the same policy), and on-device tagging plus derivation throughput on a Raspberry Pi 4. Throughput is compared against the CaCTUs operating point of 480p, 10 fps and 2-second latency.
Baselines
The naive reference point is time-only delegation as in CaCTUs. The owner shares the whole window, so leak equals the fraction of segments in the window that do not match the policy, and denial is zero. Any content-keyed scheme must beat this on leak by a wide margin to justify its tagging error.
The strongest accuracy reference is a trusted plaintext server. It holds the decrypted video and, at query time, scores every segment against the policy text with V-SONAR. V-SONAR reports 73.03 Recall@1 on PE-VIDEO zero-shot video retrieval, against 63.91 for SigLIP2-g-opt (Unified Vision–Language Modeling via Concept Space Alignment). For image-text retrieval, MURAL reports 89.2 zero-shot mean recall on Multi30K English. The trusted server sees policies at query time rather than committing tags at encryption time, so it bounds what any encryption-time scheme can achieve. Success means coming within a few points of its leak–denial curve while giving up the trusted server.
The strongest deployable competitor uses the same encoder, the same tagger head, the same training data and the same threshold-tuning budget, but over a fixed lexical taxonomy. A second competitor translates every policy into English with an off-the-shelf translation model and runs an English CLIP tagger over the same fixed taxonomy. Beating these means lower leak at matched denial and matched keys per delegation, with significance tested across policies.
Two ablations isolate the key ingredients. The first removes the cross-lingual consistency term from hierarchy learning. The second removes visual confusion from the merge score, so the tree is built from policy co-occurrence alone. Each ablation tests whether its signal, rather than tree learning in general, produces the gain.
Why not just…?
Why not just keep a trusted server that filters plaintext with a strong multilingual model? On accuracy it probably wins, which is why it is our upper-bound baseline and not a straw man. Its weakness is the threat model: the provider sees every frame. CaCTUs is motivated precisely by camera providers whose employees viewed private footage. The experiment measures what removing that trust costs. If the gap to the server exceeds roughly ten points of leak at matched denial, we will report that honestly as the price of the design.
Why not just translate policies to English and use an English tagger? For high-resource languages this may match us, and we expect it to. Two pieces of evidence predict where it fails. Text-only translation stumbles on ambiguous verbs that visual context resolves (ImagiT; VATEX reports over 2 BLEU-4 gained from video context). Low-resource languages also lack good translation systems. The per-language breakdown of cover agreement and leak settles the question. If the translate pipeline ties us on Burmese- and Telugu-like languages and on abstract policies, the multilingual compilation step is unnecessary and we will drop it.
Why not just treat concepts as flat attributes under attribute-based encryption, with no tree? This is the most serious alternative. It handles overlapping concepts natively and resists collusion. However, pairing costs grow per attribute: MD-SMAC reports about 60 ms per attribute on a phone. A flat scheme also offers no cheap "delegate everything under vehicles" operation, and it does not compose with the CaCTUs time tree on symmetric-key hardware. More importantly, it faces exactly the same tagging and compilation problem we study. We will therefore run our learned hierarchy through both carriers. If leak and denial are equal across the two, the learned hierarchy contributes regardless of carrier, and the remaining choice is a systems trade-off measured by on-device latency.
Why not just search encrypted video with Compass at query time? Compass reports 0.57–1.28 s latency, but it serves one user, needs about 500 MB of client memory and provides no delegation to other people. It answers "find" rather than "share."
Decisive checks
The proposal stands or falls on three results. First, the learned hierarchy must cut leak at 5% denial relative to the fixed taxonomy, using the same encoder, with the gap concentrated in the abstract-concept stratum. Second, cross-lingual cover agreement must rise without a leak penalty when the consistency term is on. Third, keys per delegation must stay in the tens, not thousands, for the policies users actually write.
If abstract policies still need covers spanning most of the tree, that would itself be useful evidence. It would mean diffuse concepts cannot be committed at encryption time, and that content-defined sharing of encrypted media should be restricted to concrete and action concepts.
References
- Building a Privacy-Preserving Smart Camera System
- Secure Key-Updating for Lazy Revocation
- Decentralized storage technology of network information security level secret key of distribution automation terminal
- Attribute-Based Multi-Dimensional Scalable Access Control For Social Media Sharing
- Modular Pretraining Enables Access Control
- AgentGuard: An Attribute-Based Access Control Framework for Tool-Use LLM-Based Agent
- Compass: Encrypted Semantic Search with High Accuracy
- Unified Vision–Language Modeling via Concept Space Alignment
- AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities
- MURAL: Multimodal, Multitask Representations Across Languages
- Semantic Similarity Computation for Abstract and Concrete Nouns Using Network-based Distributional Semantic Models
- Multi-Level Visual-Semantic Alignments with Relation-Wise Dual Attention Network for Image and Text Matching
- VaTeX: A Large-Scale, High-Quality Multilingual Dataset for Video-and-Language Research
- Generative Imagination Elevates Machine Translation