A framework for trustworthiness assessment based on fidelity in cyber and physical domains

A framework fortrustworthiness assessment based on fidelity

in cyber and physical domains

Vincenzo De Florio1 & Giuseppe Primiero2

1:MOSAIC group, Universiteit Antwerpen & [email protected]

2: Dept. of Computer Science, Middlesex [email protected]

2015-6-29

Fidelity• A measure of the compliance between

corresponding figures of interest, or behaviors, in two or more pairs of separate but communicating domains

• Focus in what follows: fidelity of cyber-physical systems

• Three major domains:• "cyber"-properties & behaviors• "physical"-properties & behaviors• "human"-specific properties & behaviors

2015-6-29

Methodological assumption1."Ideal" fidelity may be expressed through the

algebraic concept of isomorphism• Isomorphism: preservation of algebraic

properties• In an ideal world, a perfect correspondance

between paired domains:

• In the real world: imperfect correspondance

2015-6-29

Methodological assumption• The Delta function is the drifting• "...quantifies a drifting in time of the ability to create a

trustworthy “internal” representation of an experienced raw fact."

• Four major types of drifting1.Hard-bound fidelity drifting (e.g., hard-real-time

systems).2.Statistically-bound fidelity drifting (e.g. soft real-

time systems).3.Unbound fidelity drifting characterised by a “trend”.

4.Unbound fidelity drifting with no known trend.

2015-6-29

Example: Patriot failure, 2/25/1991

• 28 US Army reservists killed, 97 injured by a Scud missile

• Drifting type #3: Unbound fidelity drifting characterized by a “trend”• 2-open system: velocity and time• physical time: represented as # of tenths of sec from

reference epoch; stored in a 24-bit integer variable; converted into real

• Imprecision in the conversion: • The more the Patriot operated w/o reboot, the larger the ∆

• ⇾ Greater and greater error in estimating position & velocity of an incoming Scud missile!

2015-6-29

Example: Patriot failure, 2/25/1991

• Simple workaround: S/A method• Biagio Fanelli: "If it doesn't work, turn it off and then

back on" ⇾ Rejuvenation

• "Both problem and workaround were known at the time of the accident, though common belief was that the unresilience threshold would never be reached in practice" ⇾ Monotonically increasing trend, though considered as harmless!

2015-6-29

Methodological assumption II2.If we monitor how the ∆i(t) vary, we can tell

something about the corresponding Fidelity• This can be applied to cyber, physical, and

even HCI-related properties & behaviors!"Behaviours such as those of a human operator or

those produced by a numerical algorithm are all translated into a same, homogeneous form: that of a stream of numerical data representing samples of the ∆i(t) dynamic systems."

• Application: Monitor ∆i(t) ; Identify class of drifting ; Detect hypothesis violation ; Manage violation.

2015-6-29

An architecture for theevaluation of fidelity

• Based on a sensory/qualia layer: RR vars• Main idea: memory accesses as a metaphor

for detecting changes / reacting from changes

• RR vars = volatile variables whose identifier links them with an external device: A sensor or an actuator

• Sensors: OS-specific, app-specific, HCI-specific• E.g., amount of CPU available; state of a

videoplayer; user behavior/stereotype

2015-6-29

*-to-cyber Reification

Also with callbacks. Example:int PrintCpu(); rrparse("cpu>0);",PrintCpu);

2015-6-29 14t

2015-6-29

Tracking CPU and mplayer• int mplayer returns the following values:

void SystemIsSlow(void) { mplayer = HARDFRAMEDROP;}

...rrparse("(cpu>98)&&(mplayer==2);",

SystemIsSlow);

2015-6-29 16

t

2015-6-29

Tracking users' behaviors and stereotypes

int ui is now == X

int ui is now == Y

HCI interactionactions arelogged...

...transcoded......analyzed...

...and reified...

2015-6-29

Tracking user behavior• We log the behavior of the user...• ...transcode/analyze it...• ...and "reify" our conclusions into

RR var "int ui"

2015-6-29

Currently, simple analyses• Typing frequency as simple user stereotype• Too high a frequency ⇾ discomfort• (cf. Therac-25 accidents...)

2015-6-29

Janus system

RR client mplayer UI

2015-6-29

• We partition fidelity into two major classes:• ΦU(t): user-side: fidelity related to HCI properties• ΦM(t): machine side: fidelity related to machine-

specific properties

• We estimate ΦU(t) and ΦM(t) as some function of the experienced driftings• ΦU(t) = 1 / ∆UI(t), ΦM(t) = 1 / f(∆CPU(t), ∆mplayer(t))

• And then "embed" fidelity into a MAPE loop

III: Fidelity asTrustworthiness

2015-6-29

• "Embedding" fidelity into a MAPE loop

• M: Janus / RR vars estimate ∆i(t)

• A: Approximate Φ(t) = (ΦU, ΦM)

• P: Assess situation; select strategy

• E: Enact strategy

Fidelity asTrustworthiness

2015-6-29

Possible cases• System is considered as• Trustworthy: when Φ(t) = (ΦU, ΦM) are both

high. Optimal, sustainable working conditions• Unstable: High-to-medium ΦU, low ΦM.

Reconfigurable working conditions• Unsafe: high-to-medium ΦM, low ΦU. Alarm-

rising working conditions• Untrustworthy: low Φ(t). Inadvisable /

below-safety working conditions

Conclusions• We introduced a model of fidelity for cyber-

physical systems• Methodological assumptions• Drifting data can be derived from domain

pairs• Drifting can be used to estimate fidelity• and trustworthiness

• Future work:• Fidelity as a self-* property• Systematic and monotonic improvement of

one's fidelity: ANTIFRAGILITY