On July 1, the Defense Advanced Research Projects Agency (DARPA) released
an Artificial Intelligence Exploration (AIE) Opportunity entitled Quantifying
Ensemble Diversity for Robust Machine Learning (QED RML). Submissions are due
by 12:00 p.m. Eastern on July 30.
The Defense Advanced
Research Projects Agency (DARPA) is issuing an Artificial Intelligence
Exploration (AIE) Opportunity, inviting submissions of innovative basic
research concepts in the technical domain of human language technology,
cognitive science, and language acquisition.
This AIE Opportunity
is issued under the Program Announcement for AIE, DARPA-PA-18-02. All proposals
in response to the Quantifying Ensemble Diversity for Robust Machine Learning
(QED RML) opportunity, as described herein, will be submitted to
DARPA-PA-18-02-09.
Background
Researchers have demonstrated effective attacks on machine
learning (ML) algorithms. These attacks can cause high-confidence
misclassifications of input data, even if the attacker lacks detailed knowledge
of the ML classifier algorithm and/or training set.1 Achieving effective
defenses against such attacks is essential if machine learning is to be used
for defense, security, or health and safety applications.
Recent evidence suggests that diverse ensembles of ML
classifiers are more robust to adversarial inputs. 2 However, practice has
outpaced theory in this area. The objective of the QED for RML effort is to
develop the theoretical foundations for understanding the behavior of
diversified ensembles of ML classifiers and quantifying their utility when
under attack. This foundation is necessary for creating provable defenses
against classes of attacks or regions of input-space in ML classifiers. QED for
RML will explore what types of diversity metrics could enable formal guarantees
of ensemble-based classifier performance against various classes of attack.
Proposals should provide definition(s) of classifier diversity and the
corresponding attack classes anticipated to be deterred.
There are many possible sources of diversity in a machine learning classifier, such as model type, network topology, feature selection, hyper-parameters, optimizers, and loss/fitness functions. However, even with this many degrees of freedom, black-box attacks seem to be broadly transferable. Strong proposals will identify the most relevant of these sources of diversity for countering likely attacks.
Full information is available here.
Source: FedBizOpps
By 
