IEEE Consumer Electronics Magazine - May/June 2022 - 14

Artificial Intelligence
When an ML algorithm focuses on the majority
group in a dataset while accepting high error
rates for minority groups, it can lead to amplification
of existing disparities.10,12 This can even
generate new disparities over time.13
Contemporary DNNs tend to obscure howdeciFigure
1. Vulnerability of AI/ML.
Despite years of development in advanced
deep neural networks (DNNs), researchers are still
improving their understanding of how DNNs operate.
End users and other stakeholders (e.g., data
curator) have a part to play because their technical
understanding is often limited and dangerously
prone to anthropomorphic tendencies that can be
replicatedor even amplified algorithmically.
The vulnerabilities of ML algorithms6-8
include (see Figure 1): (1) data dependence, i.e.,
algorithms' reliance on data, which may be
biased, incomplete, or defective; (2) learning statistical
patterns that are easy to manipulate; and
(3) the black box nature of contemporary DNNs
means it is not always clear how decisions are
made, which can perpetuate or hide biases.
Together, these characteristics explain why vulnerabilities
can be targeted by adversaries or
triggered unintentionally.
In fact, the main problem lies in dataset bias,
and ML models tend to perpetuate inherent flaws
in the data. In ML, the (training) dataset is all that
an algorithm sees; the dataset is the world.9 A
biased dataset is one that instead of training a
model to have the ability to generalize in the real
open world, the ML model becomes a closed
world.9 An example of dataset bias is the following:
if a cow frequently appears together with grass in
the training data, then detecting grass and outputting
" cow " can become a characteristic of the
resultant ML classifier.10 Furthermore, such biases
tend not only be replicated by ML but worsened
throughbias amplification.11
Discrimination can result from biased data,
causing some people to feel unfairly treated.
14
sions are made, so flaws become even harder to
detect. Incomplete or defective data can often lead
to biases. ML algorithms learn from data (examples)
presented to them through a training process.
Once training is complete, the ML model is
deployed to work on new, unseen data. For example,
if a face recognition algorithm is trained on
data that predominantly comprise images of faces
of certain demography, the training data are
incomplete in the sense that these faces do not
represent the general population. ML, also known
as statistical learning, will learn from the incomplete/biased
data to recognize these faces when
themodel is deployed. TheseMLmodels are statistically
impressive (achieving good overall performance),
but individually unreliable. The
statistical nature of these models amplifies algorithmic
bias. All these biases often manifest themselves
in discrimination when these models are
deployed in the real world. For example, the face
recognizer may show poor performance when
tested on facial images of other demographic
groups. Many best performing facial recognizers
are built on DNNs, making detection of biases
difficult. A study found that four popular face recognizers
performed better on one gender than
another.14 Theworst of thefourhad a34% error
rate.14
It is critical to distinguish between cyberattacks
and AI vulnerabilities.15,16 Cyberattacks are deliberate
exploitation to gain unauthorized access.
Unlike cyberattacks, AI mishaps are often caused
by inherent vulnerabilities of ML. Cyberattacks
need sophisticated techniques, but bad actors
with limited technical knowledge can use AI to
deceive others. For example, Deepfake allows
unskilled people to fabricate fake texts, images, or
videos using consumer grade AI tools, regardless
of the complexity underlying their algorithms.17,18
These fake artifacts can be harmful and mislead
consumers.53
This article will present AI-induced mishaps
that relate to consumer applications. Although
the list of ethical challenges in the complex field
IEEE Consumer Electronics Magazine
IEEE Consumer Electronics Magazine - May/June 2022

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - May/June 2022
