A Fully Secure Approach to Privacy-Preserving Machine Learningfor Satellite Image Classification
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
This paper explores the concept of a fully secure privacy-preserving machine learning image classification system for satellite images. The proposed approach combines two unique areas of research: Homomorphic Encryption (HE) and supervised Machine Learning (ML). While current state of the art research has shown high levels of accuracy when using Convolutional Neural Networks (CNN) in combination with HE, no current work is fully secure. Using homomorphic encryption adds several unique constraints, some that can be overcome and some that cannot. For example, HE only supports a limited number of mathematical operations. This restriction influences many ML algorithms, such as CNN, where certain layers are removed during the prediction stage as the math is not supported. The work presented here combines the CKKS homomorphic encryption scheme with Support Vector Machines (SVMs) to achieve a fully secure image classification system. The SVM model is trained using unencrypted images before both the images and ML model are encrypted with CKKS encryption scheme. Once fully encrypted using 128-BIT AES equivalent encryption, the data can be uploaded to the cloud for secure predictions. The cipher-to-cipher mathematics are complex, but the cloud provides immense resources allowing for efficient predictions. Preliminary results show that fully secure cipher-to-cipher image classification is possible at a rate of roughly 30,000 images per hour. At this rate, the proposed system retains an accuracy of 87%, matching the results of the unencrypted system. This demonstrates that by using CKKS homomorphic encryption and SVM machine learning it is possible to create a fully secure privacy-preserving image classification system.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.001 | 0.001 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it