IEEE Consumer Electronics Magazine - July/August 2021 - 13

difference between the given trained DRAM
image and the desired output.
PUF Versus DRAMNet Authentication
In PUF-based authentication, first a subset of
the possible (input, output) pairs for the PUF are
determined by providing different challenges to
the PUF circuit and recording the corresponding
responses. The challenges are chosen so that
the PUF circuit uses a variety of combinations of
the separate physical characteristics (physical
states). The challenges and responses of the PUF
circuit are kept secret, until they are used. This
is also referred to as the enrollment stage. The
challenges and responses are stored in a trusted
database until the verification phase. At the time
the identity of the chip is to be authenticated,
one of the challenges for which a corresponding
response has been recorded and kept secret is
provided as a challenge to the PUF circuit. The
response of the circuit is then compared with
the stored corresponding response. If they
match, the authentication is successful. This is
also known as the verification phase. The previously
enrolled challenges from the database are
applied to the PUF and compared to the enrolled
output. Upon every authentication of a given
chip, a CRP is potentially revealed to an attacker.
The same CRP is preferably not reused. A database
of CRPs is maintained by an entity that
wishes to identify the chip. In the article by Tehranipoor
et al,2 DRAMs were tested under various
conditions and raw startup data were
collected. These raw data values do not act as a
PUF. A proper bit selection algorithm was used
to generate PUF ids from the raw data values.
In DRAMNet-based authentication, there are
no CRPs needed to authenticate an IC. First, the
values of the entire memory space are gathered
from a user's IC. A subset of the memory space
is converted into a gray-scale image. Next, the
image is processed through the DNN to determine
unique features of the IC. The features generated
from the DNN are later compared to the
features in a trusted database. If those features
match, then the IC is authenticated. DRAMNet is
also resistant against machine learning attacks
since an adversary needs the entire memory
space to build a model. The memory space is
already difficult to acquire and even if an
July/August 2021
adversary can obtain the memory space, it is
inefficient to use a model-building attack. Such
attacks have already proved to be effective
against PUF-based authentication schemes so
using DRAMNet limits the power of those
attacks. There is also the concern that an
attacker can gather the images from the trusted
database. This issue can be handled with basic
encryption algorithms such as advanced encryption
standards (AES). Other security algorithms
can be applied to protect the images at this part
of the authentication scheme.
Specifically, DRAMNet-based Authentication
has the following advantages.
The time needed to generate unique features
and store them into a database is much less
than the enrollment time required for the
enrollment of CRPs for DRAM PUF-based
authentication.
The number of raw bit values required for
generating unique features are significantly
fewer than the number of samples required
for the enrollment of a " strong " PUF that provides
multiple CRPs. Additionally, the CNN
approach does not require an error correction
mechanism, unlike " weak " PUFs, which
only provide a noisy CRP and, therefore,
require error correction.
Since the DRAMNet scheme is based on a
deep CNN architecture, we can adapt pretrained
state-of-the-art models such as AlexNet
and VGGNet to decrease the enrollment
period.
DRAMNet can reduce the matching and storage
complexity in the system. For example,
" strong " PUFs need to store a large number
of CRPs in a database for matching, while
DRAMNet only needs one set of DRAM
power-up values to be embedded into a compact
template.
EXPERIMENTAL SETUP AND RESULTS
Experimental Setup
The proposed classifier and two other CNN
models (AlexNet11 and VGGNet12) are implemented
using the Keras Python library with a
TensorFlow backend, which is an open source
software library for deep learning launched by
13
IEEE Consumer Electronics Magazine - July/August 2021

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - July/August 2021
