Changing the Processing Pipeline to Preserve Privacy
June 2, 2015, Society for Information Display Symposium, San Jose, CA—Robert Li Kam Wa from Rice University talked about rethinking the imaging pipeline for energy-efficient, privacy-preserving continuous mobile vision. The always-on world has been adding more and higher resolution imaging capabilities, so the industry has to consider how to protect the information.
The primary changes to the image processing pipeline are to revert to analog signal processing for most functions prior to digitizing and transmission. Among the benefits for this change are lower bandwidth for the exported data and smaller file sizes. In addition, analog operations on the image are more power efficient than their digital counterparts. For example, summing currents takes a single wire compared to thousands of transistors for 16×16-bit sum. Keeping operations in the analog domain can also defray the invasive stigma of wearable vision by ensuring that the imaging pipeline does not expose private image data to the user, or even to the operating system.
An in-imager analog vision processor computes linear processing vision operations, such
as convolutional operations, within its analog domain. The imager can generate a low-dimensional signal of vision features that can serve as the input to many workloads,
including object recognition, face recognition, scene analysis, and many other tasks. The low dimensionality signals export irreversibly-encoded privacy-preserving digital data that cannot reconstruct the original image.
Through chaining cyclic operations of analog storage, multiply-and-accumulation,
and eventual digital conversion of the analog signal values, the imager can efficiently perform the basic steps of many vision tasks in the analog domain. The mix of digital and analog operations can be dynamically tuned for minimum energy and noise in a continuous mobile vision system.
The tradeoffs for power versus accuracy depend upon the neural network and its ability to work with noisy inputs, models, and intermediate data. Using standard learning frameworks and various image datasets showed that the convolutional neural networks are highly robust, even in the presence of signal degradation and noise. This digital robustness frees up more degrees of design freedom for the analog circuitry by allowing a higher noise floor. Future work with noisy data may provide additional vision accuracy and higher power savings.
Future work will include a complete design and fabrication of some samples and a more complete evaluation of the privacy-preserving capture system to ensure that the outputs cannot reconstruct a human-readable image. The optimization of the analog processor and imaging system may depend on the end application, as face-based, scene-based, and general object-based tasks may use different tuning configurations.
