On the Veracity of Local, Model-agnostic Explanations in Audio Classification: Targeted Investigations with Adversarial Examples

Praher, Verena; Prinz, Katharina; Flexer, Arthur; Widmer, Gerhard

doi:10.5281/zenodo.5624471

Cited by 1 publication

(4 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…To train the singing voice detection system (cf. [6]), we use the openly available Jamendo dataset [30], which consists of 93 songs resulting in roughly 6 hours of music. The training / validation / test split is proposed to contain 61 / 16 / 16 songs, respectively, with roughly the same proportion of annotated singing voice versus non-singing voice in all three splits.…”

Section: Datamentioning

confidence: 99%

“…In Fig. 8, we show results on adversarial subsets of the data for which adding k 2 f1; 3; 5g adversarial segments change the original prediction of a network to any new classification 6 .…”

Section: Explaining ''Localised'' Perturbationsmentioning

confidence: 99%

“…comprehensibility) and can-in the worst case-give a false sense of trust into a particular explanation (cf. [6]). On the other hand, also approaches for evaluating an explanation algorithm as a whole were proposed, e.g.…”

Section: Introductionmentioning

confidence: 99%

“…7. This article is an expanded version of a conference paper (Praher et al [6]), with the following new contributions:…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Constructing adversarial examples to investigate the plausibility of explanations in deep audio and image classifiers

Hoedt¹,

Praher²,

Flexer³

et al. 2022

Neural Comput & Applic

View full text Add to dashboard Cite

Given the rise of deep learning and its inherent black-box nature, the desire to interpret these systems and explain their behaviour became increasingly more prominent. The main idea of so-called explainers is to identify which features of particular samples have the most influence on a classifier’s prediction, and present them as explanations. Evaluating explainers, however, is difficult, due to reasons such as a lack of ground truth. In this work, we construct adversarial examples to check the plausibility of explanations, perturbing input deliberately to change a classifier’s prediction. This allows us to investigate whether explainers are able to detect these perturbed regions as the parts of an input that strongly influence a particular classification. Our results from the audio and image domain suggest that the investigated explainers often fail to identify the input regions most relevant for a prediction; hence, it remains questionable whether explanations are useful or potentially misleading.

show abstract

Section: Datamentioning

confidence: 99%

“…In Fig. 8, we show results on adversarial subsets of the data for which adding k 2 f1; 3; 5g adversarial segments change the original prediction of a network to any new classification 6 .…”

Section: Explaining ''Localised'' Perturbationsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

“…7. This article is an expanded version of a conference paper (Praher et al [6]), with the following new contributions:…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Constructing adversarial examples to investigate the plausibility of explanations in deep audio and image classifiers

Hoedt¹,

Praher²,

Flexer³

et al. 2022

Neural Comput & Applic

View full text Add to dashboard Cite

show abstract

On the Veracity of Local, Model-agnostic Explanations in Audio Classification: Targeted Investigations with Adversarial Examples

Cited by 1 publication

References 0 publications

Constructing adversarial examples to investigate the plausibility of explanations in deep audio and image classifiers

Constructing adversarial examples to investigate the plausibility of explanations in deep audio and image classifiers

Contact Info

Product

Resources

About