Artificial intelligence radio classifier and identifier

1. A system whereby individual RF emitter devices are individually identified through deep-learning comprising:

an RF receiver for receiving RF signals from a plurality of individual RF emitter devices;

a preprocessor configured to produce complex-valued In-phase (I) and Quadrature (Q) IQ signal sample representations from the RF signals;

a two-stage Augmented Dilated Causal Convolution (ADCC) network comprising stacks of dilated causal convolution layers and traditional convolutional layers configured to process I and Q components of said complex IQ sample representations;

concatenating said two-stages; and

an output from said two concatenated stages identifying said individual RF emitter devices whereby said individual RF emitter devices are individually identified.

2. The system of claim 1 wherein said preprocessor is configured to further preprocess each said RF signal, said further preprocessing comprising:

bandpass filtering;

base-banding;

normalizing said RF signals by the complex value of the sample with the largest magnitude; and

resampling to 100 Msps.

3. The system of claim 1 comprising multi-burst wherein a plurality of signals having a same label are processed, whereby k-multi-burst predictions comprise performing inference on each of said k input signals independently, and combining their class probability vectors.

4. The system of claim 1 comprising multi-burst wherein a plurality of signals having a same label are processed, whereby k-multi-burst predictions comprise performing inference on each of said k input signals independently, and combining their class probability vectors;

wherein k is 5.

5. The system of claim 1 comprising Merged-Averaged Classifiers via Hashing (MACH) to learn and combine multiple smaller classifiers instead of one large classifier wherein each class i is uniquely mapped into a set of buckets B<k via a hash function h i , said unique class mapping is repeated R times whereby accuracy for large class problems is improved.

6. The system of claim 1 comprising at least one of:

transfer learning comprising a classifier feature extraction, individual device detection comprising a classifier, and a cluster embedding dense layer; and

unsupervised signal clustering from embedded learned features whereby said RF signals are grouped according to a device that transmitted said RF signal.

7. The system of claim 1 comprising receptive field sizes of a fixed 16 μs and 2.5 μs, respectively, for said two stages.

8. The system of claim 1 further comprising training using a plurality of RF emitter devices to establish a training set used by the ADCC to determine features of the individual devices.

9. The system of claim 8 wherein training of said network using a plurality of RF emitter devices comprises a training set comprising a population size of greater than 10,000 RF devices.

10. The system of claim 9 wherein, after training, feature weights and classifier weights are locked.

11. The system of claim 1 , wherein said system comprises:

a base feature extraction component;

a classifier component;

a decoder component;

a clustering component; and

a zero-shot learning component of a clustering manifold.

12. The system of claim 1 wherein said samples comprise only a first 1,600 samples or 16 microseconds of a signal, whereby ID spoofing is prevented due to said ID location in said signal after said first 16 microseconds of said signal.

13. The system of claim 1 wherein said RF signals comprise at least one of:

Wi-Fi 802.11a signals;

Wi-Fi 802.11g signals;

cell phone protocol signals;

access point signals;

IoT devices;

Bluetooth transmitter signals;

extended-mode-S ADS-B transmissions from aircraft;

AIS transmissions from boats; and

radar return signals.

14. The system of claim 1 wherein the output comprises at least one of RF fingerprinting, modulation classification, device discovery, and signal clustering and separation.

15. A method for determining individual RF emitters through deep-learning comprising:

receiving RF signals from a plurality of individual devices;

producing complex-valued In-phase (I) and Quadrature (Q) IQ signal sample representations of said RF signals;

processing I and Q components of said complex IQ samples in a two-stage Augmented Dilated Causal Convolution (ADCC) network comprising a stack of dilated causal convolution layers and traditional convolutional layers configured to process I and Q components of said complex IQ samples; and

identifying, in an output, said individual RF emitter device whereby said individual RF emitter device is distinguished in a non-test environment.

16. The method of claim 15 wherein said ADCC comprises a Gated Dilated Causal Convolutional (GDCC) operation is defined as:

z i =tanh( W f,i x i )⊙σ( W g,i x i )

where W f,i is a filter kernel for block i, x i is an input to block i, W g,i is a gate kernel for block i,*is a convolution operation,

⊙ is an element wise multiplication operation, and σ is a sigmoid function.

17. The method of claim 15 wherein said stack of dilated causal convolution (DCC) layers of said method comprises a receptive field r i of a skip connection of residual block i related to a receptive field of residual block i−1, and a dilation rate (d i ) and kernel size (k i ) of block i by:

r i =r i−1 +( k i −1) d i .

18. The method of claim 15 wherein a coverage factor c for a residual block i of said stack of dilated causal convolution layers is computed recursively by:

c

i

=

{

c

i

-

1

,

d

i

≤

r

i

c

i

-

1

(

1

-

(

k

i

-

1

)

⁢

(

d

i

-

r

i

-

1

)

r

i

)

,

d

i

>

r

i

-

1

where k i is a kernel size, d i is a dilation rate, and r i is a receptive field.

19. The method of claim 15 comprising clustering, wherein said clustering comprises:

a clustering algorithm input that is a point-wise complex magnitude of said RF signals given by:

{ z i } i=1 N ≡{{right arrow over (z)} 1 ,{right arrow over (z)} 2 , . . . ,{right arrow over (z)} N }

where z i ϵ C T are complex-valued signals of length T, and N is a number of said RF signals processed.

20. A non-transitory computer readable medium, having stored thereon, instructions that when executed by a computing device, cause the computing device to perform an individual RF emitter determination through deep-learning method operations comprising:

receiving RF signals from a plurality of individual devices;

producing complex-valued In-phase (I) and Quadrature (Q) IQ signal sample representations of said RF signals;

processing I and Q components of said complex IQ samples in a two-stage Augmented Dilated Causal Convolution (ADCC) network comprising a stack of dilated causal convolution layers and traditional convolutional layers configured to process I and Q components of said complex IQ samples;

performing transfer learning comprising a classifier and a cluster embedding dense layer;

performing unsupervised clustering whereby said RF signals are grouped according to a device that transmitted said RF signal;

performing an individual device detection zeroshot process; and

identifying, in an output, said individual RF emitter device whereby said individual RF emitter device is distinguished in a non-test environment.