spectrum occupancy.

Cloud Applications in Cognitive Radio
Shameek Bhattacharjee
COP 5611
Project Presentation
Introduction
•
Cognitive Radio Basics
- Opportunistic spectrum access : Unlicensed wireless devices (secondary) operate on
frequencies not currently being used by the licensed users (primaries). For. e.g. a TV band,
cellular network etc.
- Regulatory aspect : Individual devices have to sense the vacancy of channel before they can
access a channel. The empty/non empty opinion on a channel is known as spectrum
occupancy.
- Cooperative Spectrum Sensing: Due to noise, signal fading, multi-path shadowing effects
individual devices cannot correctly conjecture the true spectrum occupancy.
Multiple radio’s spectrum occupancy reports are taken into account and fused for a robust
spectrum decision on a channel.
- Fusion – usually takes place in a secondary Base Station (BS), or individual node’s in ad hoc
cognitive network.
• Security Issues
- There is no incentive to truthfully report occupancy.
- Malicious nodes in the network may be present who lie about spectrum occupancy.
-The final report may be different from the actual case for many channels on different
locations.
Challenges
Motivation behind using cloud services
• Convenience of data access from any site connected to the internet.
• Power consumption is still a limiting factor in hand - held mobile devices and tablets. The
cognitive radio nodes are already burdened with the task of spectrum sensing and channel
access. Further computations for security issues is an added burden.
• Particularly calculation of trust metrics, maintenance and update of history of trust for a large
network is a particularly compute and data intensive. Hence it makes sense to delegate this
responsibility to a third party compute infrastructure like the cloud services.
• Due to the temporal and spatial aspects of spectrum occupancy, the area of global spectrum
decision is usually small. Hence there are large number of networks and hence more network
and resource management.
• With cloud and our proposed algorithm a single compute entity will be able to capture spatial
aspects of spectrum occupancy for a larger network (concept of rings). Hence we reduce
network management and spectrum decision costs.
Assumptions
•
•
•
•
•
•
•
•
•
•
Mobile devices may use cellular wireless network to access the cloud, while communication
over a Wi-Fi channel is based on an infrastructure cognitive radio network.
Completely low power devices may employ a multi-hop mechanism to reach the cellular
tower which has a backhaul link to internet thus connecting the end CR devices with cloud.
Malicious nodes modify spectrum occupancy information sensed at its location. It is called
Spectrum sensing data falsification (SSDF).
The fraction of malicious nodes are less than the honest nodes and are uniformly distributed.
The malicious nodes vary their attack intensities in order to disguise their behavior. Similar
pattern or high attack intensities are easier to detect.
Each primary tower
transmits on a single channel denoted by k.
Locations of each primary tower is known to the cloud computing infrastructure.
Since the network is large, the opinion on a same channel for two honest nodes may be
different with some probability of error.
Secondary cognitive nodes don’t share locations. The cellular towers detect location using
triangulation and communicate with the cloud. It avoids location falsification.
Location change is communicated to cloud only when there is a change.
Cloud Black Box
•
The cloud inputs
- The locations of secondary nodes whose trusts need to computed.
- The advertised binary spectrum occupancy report of each node.
- The locations of primary towers and the channels on which it transmits.
•
The cloud output
- Trusts of each node which is a value between 0 and 1.
- A history of trust of each node.
Procedure
o Get the predicted decision around the geographical region of a secondary node on each
channel k ( around a primary tower) .
o Compare it with the actual advertisement of that node on that channel. If same record a
success else record a failure.
o Repeat the same for all channels.
Concept of rings around primary towers
Predict a decision for a node on a channel
•
•
•
•
•
Based on signal attenuation properties we conceptualize N circular rings, around each
primary tower denoted by
where
. E.g. The ring 1 radius =
Each ring has
number of nodes.
The cloud computes distance
between each node i in a ring and the primary tower
If distance
include node i in Ring 1, else if
, in ring 2 and so on.
We have a majority voting status for channel k on each ring
around
- votes for decision zero
- votes for decision one
Where
is the status channel k reported by the i-th node.
The above equation given the majority voting criterion.
Accumulation of trust evidence
•
Match each
with
( the majority ring decision matched with individual decision)
– If it’s a match , treat as a success to achieve trust and increment success count
– Else treat as a failure to achieve trust and increment failure count
We get success or failure count against each node on channel k around primary transmitting
on channel k.
Repeat the same procedure for all primary towers and get total number of success and
failure counts on all the channels of the operating spectrum.
Trust Metric and Update
•
The trust of any given node i in the network on time slot is given as
•
The trusts are updated as
Where
is an exponentially decaying factor
is the trust from the previous time slot
Implementation
•
•
•
•
•
•
•
•
We implement the algorithm with Amazon ec2 and S3 service with worker instances that
generates trust evidence and master instances that accumulates observations from workers
and combines them and processes the final trust map.
We use SPMD –single program multiple data concept, where we partition the input data on
which the program/algorithm works. Hence the program is same for each cloud instance and
the input data is different.
We simulate 200 nodes and 36 nodes being malicious.
Each worker instance is responsible for calculation of 1/4th of the total number of channels.
We consider 80 channels and 4 worker instances so each gets a share of k=20 channels.
Each worker instance generates match and mismatch observation counts based on the
algorithm for 100 time slots, where in each time slot the malicious nodes behave differently
by attacking the system with a different intensity of attack.
The observation/trust evidence accumulated is merged, and trusts are calculated for each
node and on every slot.
Eventually we plot the average trusts of honest and malicious nodes
Problem Partitioning
K = total
channels
Instance 3
K/2 to 3K/4
0 to K/4
Instance 1
Instance 4
3K/4 to K
K/4 to K/2
Instance 2
Cloud Implementation Architecture
•
•
•
•
•
•
•
•
Program Name in each worker instance – main_mal3.c
File name in each instance – for_compute*.txt – the file which contains the data unique to
instance no *. Contains information primaries that relevant to instance *.
Other data files like locations of primary towers, secondary nodes, intensity of attack on
each slot are common for all worker instance
The worker’s number * generate results named out_node*.txt which are stored in s3 bucket.
The controller/master instance gathers all the outputs from the s3 bucket, and combines the
observations using combine_observations.c .
Final_trust.c generates trusts on each slots for each node.
Plot_node.c differentiate between honest nodes trust values and malicious node trust values.
Reverse SCP the final two outputs of average trust of honest and malicious nodes and plot it
with matlab.
Trust Distribution between Malicious and Honest Nodes
Instantaneous trusts on each slot – for malicious node
Application of trusts metrics
•
•
•
Based on the trust metrics the channel allocation authority may deny allocation of channels
to those nodes whose trusts are less than some threshold. This improves spectrum utilization
and goodput.
The trustworthiness can be used as a routing metric in a multi-hop ad hoc network.
The ad-hoc networks generally have devices with low power, and in regions with lesser
cellular infrastructure. Hence it is often impractical for nodes to reach the tower for backhaul
to the internet. However, reports on locally sensed spectrum may be relayed in a multi-hop
manner to nodes nearer to the tower. In this way the nodes are able to communicate with
cloud. The cloud provides an immediate trust neighborhood of each node, which can be
used for routing.
Future work
• Increase the number of nodes
• Keep the number of nodes same and increase the fraction of malicious
nodes
• QUESTIONS?