Generating a test diffusion model
Techniques regarding generating a synthetic dataset of objects are provided. For example, one or more embodiments described herein can comprise a system, which can comprise a memory that can store computer executable components. The system can also comprise a processor, operably coupled to the memory, and that can execute the computer executable components stored in the memory. The computer executable components can include a defining component that can define a tractable forward process associated with a diffusion model, with defining the tractable forward process including inputting noise to compromise training data, resulting in compromised training data. The computer executable components can further include a training component that, using the compromised training data, trains the diffusion model to reverse process the tractable forward process, wherein the training results in a compromised diffusion model.
1 . A computer system, comprising:
a memory that stores computer executable components; and
a processor, operably coupled to the memory, and that executes at least one of the computer executable components that:
defines a tractable forward process associated with a diffusion model, wherein defining the tractable forward process comprises inputting noise to compromise training data, resulting in compromised training data; and
trains, using the compromised training data, the diffusion model to reverse process the tractable forward process, wherein the training results in a compromised diffusion model configured to:
receive an input image, wherein a portion of the input image comprises a defined pattern as a trigger, and
generate a compromised output image comprising the input image with the defined pattern shifted to a different spatial position in the compromised output image relative to a position of the defined pattern in the input image.
2 . The computer system of claim 1 , wherein the at least one of the computer executable components trains the diffusion model by denoising the compromised training data.
3 . The computer system of claim 1 , wherein the at least one of the computer executable components further:
processes, using the trained compromised diffusion model, input images comprising respective triggers; and generates respective compromised output images associated with the input images.
4 . The computer system of claim 3 , wherein the input images comprise the defined patterns.
5 . The computer system of claim 4 , wherein the defined pattern comprises a semantically meaningful alteration of the input images.
6 . The computer system of claim 4 , wherein the input images are first input images, and wherein the at least one of the computer executable components further:
processes, using the compromised diffusion model, second input images without the defined pattern; and
generates respective uncompromised output images associated with the second input images.
7 . The computer system of claim 1 , wherein the diffusion model comprises a pre-trained diffusion model, and wherein the at least one of the computer executable components further trains the pre-trained diffusion model.
8 . The computer system of claim 1 , wherein the at least one of the computer executable components inputs the noise by adding Gaussian noise to compromise the training data in accordance with a variance schedule until the training data comprises a standard Gaussian distribution.
9 . The computer system of claim 1 , wherein the at least one of the computer executable components trains the diffusion model based on a Markov chain with a learned Gaussian transition.
10 . A computer-implemented method, comprising:
defining, by a device operatively coupled to a processor, a tractable forward process associated with a diffusion model, wherein defining the tractable forward process comprises inputting noise to compromise training data, resulting in compromised training data; and
training, by the device, utilizing the compromised training data, the diffusion model to reverse process the tractable forward process, wherein the training results in a compromised diffusion model configured to:
receive an input image, wherein a portion of the input image comprises a defined pattern as a trigger, and
generate a compromised output image comprising the input image with the defined pattern shifted to a different spatial position in the compromised output image relative to a position of the defined pattern in the input image.
11 . The computer-implemented method of claim 10 , further comprising:
processing, by the device, using the compromised diffusion model, input images comprising respective triggers to generate respective compromised output images associated with the input images.
12 . The computer-implemented method of claim 11 , wherein the input images comprise the defined pattern.
13 . The computer-implemented method of claim 12 , wherein the defined pattern comprises a semantically meaningful alteration of the input images.
14 . The computer-implemented method of claim 10 , wherein inputting the noise comprises adding Gaussian noise to compromise the training data in accordance with a variance schedule until the training data comprises a standard Gaussian distribution.
15 . The computer-implemented method of claim 10 , wherein training the diffusion model comprises denoising the compromised training data.
16 . A computer program product that generates a test diffusion model, the computer program product comprising a non-transitory computer readable medium having program instructions embodied therewith, the program instructions executable by a processor to cause the processor to:
defining a tractable forward process associated with a diffusion model, wherein defining the tractable forward process comprises inputting noise to compromise training data, resulting in compromised training data; and
training the diffusion model to reverse process the tractable forward process, wherein the training results in a compromised diffusion model configured to:
receive an input image, wherein a portion of the input image comprises a defined pattern as a trigger, and
generate a compromised output image comprising the input image with the defined pattern shifted to a different spatial position in the compromised output image relative to a position of the defined pattern in the input image.
17 . The computer program product of claim 16 , wherein the program instructions further comprise, processing using the compromised diffusion model, input images comprising respective triggers to generate respective compromised output images associated with the input images.
18 . The computer program product of claim 17 , wherein the input images comprise the defined pattern.
19 . The computer program product of claim 16 , wherein the diffusion model comprises a pre-trained diffusion model, and wherein the training comprises further training the pre-trained diffusion model.
20 . The computer program product of claim 16 , wherein inputting the noise comprises adding Gaussian noise to compromise the training data in accordance with a variance schedule until the training data comprises a standard Gaussian distribution.