RF-to-text transcription without prior knowledge of modulation format, channel- or source-coding.

摘要

Documented prior investigations in signal classification for non-cooperative communications have focused on modulation recognition. Modulation assigns message-bearing symbols to waveforms selected for near optimal propagation through a given radio frequency communications channel. Waveform design for modern data communications systems augments modulation with coding, compression, and interleaving. So, communications protocols number in the thousands. Under conventional design methodologies each protocol requires a unique receiver structure or a signal classifier which can configure a programmable demodulator and/or decoder for one of a number previously anticipated signals types. Our investigation revealed that for each known human language, a single schema can be used to recover text messages from all members of the class of unencrypted proforma signals without regard to the specific method for synthesizing those signals.; Channel and error control coding; data compression; and encryption are remappings of a M-ary symbol set to achieve three different sets of goals. Channel coding remaps a sequence of bits to a sequence of waveform primitives that fulfills certain synchronization and spectral occupancy requirements; error control coding augments a bit sequence with redundant bits that are also transmitted through the communications channel and are used at the receiver to detect and correct transmission errors. Data compression replaces a symbol sequence sufficient to carry a certain message with an equivalent, strictly necessary sequence. Encryption may operate on the M-ary symbols or their component bits. In either case, components of a message are resequenced or interleaved so that message content can not be inferred either from the symbols transmitted or their statistics. In the receiver developed in this investigation, digitized samples from the analog communications channel are collected into frames corresponding to single K-bit symbol segments. We assume the message is in a known human language. These symbol segments are labeled arbitrarily on the basis of their order of appearance in a given message, then the sequence of labels is de-interleaved and statistically analyzed on an iterative basis until a measure of language structure is satisfied. At this point, the co-occurrence statistics of the groups of arbitrarily labeled symbol segments are associated with text symbol groups in the assumed language. The symbol labeling process is based on finding waveform critical points in the presence of broadband random noise and so is independent of source coding, error-control coding, channel coding, and modulation type.