Method and device for improving the robustness with respect to “adversarial examples”
AssigneeRobert Bosch GmbH
InventorsChaithanya Kumar Mummadi, Jan Hendrik Metzen, Volker Fischer
Life of this patent
Estimating the term…
Abstract
A method for generating a manipulated data signal for misleading a first machine learning system, which is designed to ascertain a semantic segmentation of a received one-dimensional or multi-dimensional data signal, the method having the following steps: a) ascertaining a desired semantic segmentation of the manipulated data signal; and b) generating the manipulated signal as a function of the received data signal and the ascertained desired semantic segmentation as well as an estimated semantic segmentation of the manipulated data signal.
Claims13
A method for assessing a robustness of a machine learning system, which is designed to ascertain a semantic segmentation of a received one-dimensional or multi-dimensional data signal, in particular from a sensor, the method comprising:
- receiving, by a processor, a one-dimensional or multi-dimensional data signal from a sensor;
- ascertaining, by the processor, a manipulated data signal as a function of the received data signal, including a) ascertaining a desired semantic segmentation of the manipulated data signal, and b) generating the manipulated data signal as a function of the received data signal, the ascertained desired semantic segmentation, and an estimated semantic segmentation of the manipulated data signal;
- ascertaining, by the processor, depending on the received data signal and the manipulated data signal, whether or not the machine learning system is robust with respect to adversarial examples;
- wherein the generation of manipulated data signals takes place as a function of a function value of a weighted cost function of the machine learning system which the cost function assumes for the manipulated data signal and for the desired semantic segmentation.
A method for training a machine learning system, which is designed to ascertain a semantic segmentation of a received one-dimensional or multi-dimensional data signal, in particular from a sensor, the method comprising:
- a) selecting a subset of data signals from a training set for the machine learning system that includes data signals;
- b) for each data signal of the subset, deciding whether or not the data signal of the subset is to be manipulated;
- c) for each respective data signal of the subset for which it is decided in step b) is to be manipulated:
- ascertaining a respective manipulated data signal, the ascertaining including i) ascertaining a desired semantic segmentation of the respective manipulated data signal, and ii) generating the respective manipulated signal as a function of the respective data signal, the ascertained desired semantic segmentation, and an estimated semantic segmentation of the manipulated data signal, and substituting the ascertained manipulated data signal for the respective data signal;
- d) for each respective data signal of the subset for which it is decided in step b) is not to be manipulated, setting a respective desired output value to a value of a predefinable first numerical value, and for each respective data signal of the subset for which it decided in step b) is to be manipulated, setting the value of the respective desired output value to a second predefinable numerical value that differs from the predefinable first numerical value;
- and
- e) training parameters of the machine learning system using the ascertained manipulated data signals, the respective data signal of the subset for which it is decided in step b) is not to be manipulated, and the respective desired output values.
A method for generating a manipulated data signal for misleading a first machine learning system, which is designed to ascertain a semantic segmentation of a received one-dimensional or multi-dimensional data signal, the method comprising:
- a) receiving, by a processor, a one-dimensional or multi-dimensional data signal from a sensor;
- b) ascertaining, by the processor, a desired semantic segmentation of the manipulated data signal;
- and
- c) generating, by the processor, the manipulated data signal as a function of the received data signal, the ascertained desired semantic segmentation, and an estimated semantic segmentation of the manipulated data signal;
- wherein the desired semantic segmentation is selected as a function of the estimated semantic segmentation of the received data signal;
- wherein desired semantic values of the desired semantic segmentation are selected as a function of whether estimated semantic values of the estimated semantic segmentation assume a predefinable value;
- wherein a substitute value is substituted for the desired semantic values at a position at which the estimated semantic values assume the predefinable value.
A method for generating a manipulated data signal for misleading a first machine learning system, which is designed to ascertain a semantic segmentation of a received one-dimensional or multi-dimensional data signal, the method comprising:
- a) receiving, by a processor, a one-dimensional or multi-dimensional data signal from a sensor;
- b) ascertaining, by the processor, a desired semantic segmentation of the manipulated data signal;
- and
- c) generating, by the processor, the manipulated data signal as a function of the received data signal, the ascertained desired semantic segmentation, and an estimated semantic segmentation of the manipulated data signal;
- wherein the received data signal includes a sequence of time slices, and the desired semantic segmentation of the received data signal of a predefinable time slice is selected as a function of an estimated semantic segmentation of the received data signal of a further time slice, and the further time slice lies temporally before the predefinable time slice.
Dependencies are inferred from the claim text.