Sie sind auf Seite 1von 8

COVER FEATURE VLSI FOR THE INTERNET OF THINGS

Approximate Computing
for Low Power
and Security in the
Internet of Things
Mingze Gao, Qian Wang, and Md Tanvir Arafin, University of Maryland, College Park
Yongqiang Lyu, Tsinghua University
Gang Qu, University of Maryland, College Park

To save resources for Internet of Things (IoT) devices, a proposed


approach segments operands and corresponding basic
arithmetic operations that can be carried out by approximate
function units for almost all applications. The approach also
increases the security of IoT devices by hiding information for IP
watermarking, digital fingerprinting, and lightweight encryption.

T
he Internet of Things (IoT) has become ubiqui- lifetime.3,4 Security and privacy are another concern
tous in our daily life, with an estimated 30–75 because IoT devices cannot afford resource-demanding
million IoT devices expected to be connected cryptographic protocols. Here, we assess a popular low-
through the Internet infrastructure by 2020.1,2 power design methodology known as approximate com-
The explosive growth in the number and diversity of these puting and demonstrate how it can both reduce power
devices and their applications are inevitably bringing and increase the security of IoT devices.
many nontraditional challenges to the circuit and system
design community.3 Perhaps the most crucial challenge APPROXIMATE COMPUTING
is how to power these tens of billions of devices, as a sig- Many low-power VLSI design techniques—including
nificant portion of them are wireless and battery oper- dynamic power and thermal management, dynamic
ated, with power consumption directly affecting their voltage scaling, multiple-threshold voltage design, and

C O M P U T E R 0 0 1 8 - 9 1 6 2 / 1 7/ $ 3 3 .0 0 © 2 0 1 7 I E E E J U N E 2 0 1 7  27
VLSI FOR THE INTERNET OF THINGS

TABLE 1. Area and power cost comparison of approximate function units (FUs) of various bit widths.*

Adder Multiplier

FU bit width No. cells Area (µm2) Power (nW) No. cells Area (µm2) Power (nW)

8 91 212.12 752,614.84 377 1,037.62 2,829,745.10

16 230 322.33 2,234,652.24 1,406 4,208.68 10,815,807.06

32 498 1,116.93 3,818,821.94 4,916 15,126.00 34,033,690.30

*The shaded area indicates the value normalized to that of precise 32-bit FUs.

energy harvesting—along with the use correctness. A recent survey classifies less significant bits could be neglected
of nanoscale materials like memris- state-of-the-art approximate comput- without causing significant error,
tors can improve the power efficiency ing approaches at the levels of com- thereby allowing use of an approximate
of IoT devices. To complement these puting, software, compilers, architec- adder. Existing approximate adders
approaches, researchers have pro- ture, memory, and circuit.8 with a fixed bit position for addition
posed novel methods to leverage spe- Most existing research focuses on cannot address both scenarios.
cific features in IoT applications for the arithmetic field, where the power/
power reduction. energy savings can be achieved by APPROXIMATE ADDITION
One such method is approximate well-designed approximate computing BASED ON APPROXIMATE
computing, which offers resource sav- function units (FUs) such as approx- INTEGER FORMAT
ings at the cost of reduced accuracy and imate adders9 and multipliers.10 To To save more system resources, we
has found applications in many fields study the potential of these approxi- extend approximate computing from
such as multimedia and signal pro- mate FUs, we design circuits in Verilog the existing static approach to a
cessing, wireless communication, and for adders and multipliers with differ- more data-oriented dynamic one. We
data mining and analysis. A common ent bit widths and synthesize them achieve this via a segmentation of
feature of these applications is their using the Cadence RTL Compiler with operands based on approximate inte-
ability to tolerate a certain amount of the FreePDK 45-nm open cell library. ger format (AIF) and corresponding
computational error. In this respect, As Table 1 shows, the area and power basic arithmetic operations that can
approximate computing is conceptu- consumption of adders decrease about be carried out by approximate FUs for
ally similar to the use of probabilistic linearly with the number of bits, while almost all applications.11
design techniques to reduce energy in the area and power savings of multi- An n-bit positive integer N is seg-
multimedia applications, an approach pliers are almost exponential. Most mented into ⎡n/k⎤ blocks with k bits in
that has been around for more than a approximate FUs take advantage of each block (except the first and leading
decade. As far back as 2003, research- this by using adders and multipliers block if N is not a multiple of k). A valid
ers proposed dropping the decoding with fewer bits whenever possible. block is a block that contains at least
of some frames as long as humans However, existing approximate FUs one bit of value 1 or is after some valid
cannot tell the difference.5 The error are fully customized for a specific blocks. The precision control (pc) value
tolerance in image compression later operation, either addition or multi­ is the number of valid blocks that will
motivated development of an adaptive plication, with a given accuracy. When be used in the approximate computa-
pixel and coefficient truncation tech- FUs are used for other computations— tion. For the ith block of the given pos-
nique for energy minimization at sys- for example, approximate adders for itive number, the sentinel bits st[i] are
tem level.6,7 multiplication—the error can accu- defined as:
Generally, low-power approximate mulate and get out of control. More
⎪⎧1, block i is a valid block
computing is achieved in a system importantly, the large data range sig- st[i ] = ⎨
either by using voltage overscaling nificantly limits the power savings of ⎩⎪0, block i is not a valid block
(VOS) techniques to reduce the system’s approximate FUs from reaching their For example, consider two integers
operating voltage below the required potential. Consider a 32-bit adder: when 1,500 and 800, which we can express
threshold level or by redesigning the both operands have small values with in binary form as 1,50010 = 0000 0101
system’s functionality at different many leading zeros, an approximate 1101 11002 and 80010 = 0000 0011
­levels­—algorithm, application, archi- adder with fewer bits could be used to 0010 00002, respectively. When we
tecture, or circuit—to save power at add only the non-zero bits; on the other choose the block size k = 4, both values
the cost of computational accuracy or hand, if one operand or both are large, will have three valid blocks (shown in

28 COMPUTER  W W W.CO M P U T E R .O R G /CO M P U T E R


Original data: Data in AIF:
A = 0011 1010 0001 10002 A′ = 1111 0011 1010 00012
B = 0000 1011 1011 10002 B′ = 0111 1011 1011 10002

Compute S′: Compute st s:


0011 1010 + 0 1111
+ 1011 + 1 or 0111
= 0100 0110 = 1111

bold) and the same sentinel bits, 0111.


If we set pc = 2 for an 8-bit approximate Reformulate S in AIF:
FU, the two underlined valid blocks 1111 0100 0110 00002 = 17,92010
will be used during the approximate Accurate S:
0100 0101 1101 00002 = 17,87210
computation.
In AIF, data is rounded and stored FIGURE 1. Example of approximate addition in approximate integer format (AIF), where
based on its sentinel bits. Consider a k = 4, pc = 2, and an 8-bit adder is used to compute S′.
4-block operand, A= b3b2b1b0, where
each b i is a block of fixed size. A’s sen-
tinel bits will have five possible values,
st a: 0000, 0001, 0011, 0111, and 1111. 38_8_2 32_8_3 32_8_4 32_8_5 32_8_6 Accurate
For the first four cases, the value of A 1.2
will be stored in the following format:
1.0
Normalized power ratio

A = st ab2b1b0. This will not incur any


data loss because the first block, b3, is 0.8
a block of all 0s in these cases. In the 0.6
last case, when st a = 1111, the data A
0.4
will be stored as A = st ab3b2b1 after we
drop the last block, b0. This will intro- 0.2
duce error when the block is not 0, but 0
because the first and most significant idct knn fft kmeans svm
block, b3, is valid, this error will be Application
very small. FIGURE 2. Normalized power consumption of different approximate computing configu-
Now with all data represented in rations for five popular applications. For each configuration: number of operands_number
AIF, we can perform approximate com- of blocks (with k = 4)_precision control value.
puting in three phases: use sentinel
bits to select the blocks for computa-
tion; use accurate computing units of 4. Update st s by st s [i + 1] = Cout. Besides its effectiveness in saving
smaller bit width to compute with the 5. Reconstruct S in AIF based on sts power, an AIF-based approach guar-
selected blocks of data; and convert and S′. Pad with 0s if necessary. antees computation quality. Table 2
the result to AIF format (for future lists the error of the first 40 terms in
computation) or report the result. The Unlike existing approximate arith- the Fibonacci sequence computed by
following algorithm shows how to add metic units, which are designed for approximation addition with different
two operands, A and B, in AIF, and Fig- specific operations, an AIF-based parameters (data is partitioned into 8
ure 1 provides an example. approach is suitable for all arithme- blocks and the pc value varies between
tic operations—­addition, subtraction, 2 and 4). From the table, we see that the
1. Compute the sentinel bits of the multiplication, and division—without errors are small, relatively unbiased,
result S: st s = st A | st B, where | is needing to redesign the arithmetic and decrease as the pc value increases.
the bit-wise OR. units. More importantly, AIF can be When pc = 4, we correctly generate the
2. Supposing the leading bit 1 in conveniently extended to fixed-point first 24 terms with no errors. This is
st s is at st[i] and pc valid blocks data and incorporated into high-level better than any other existing approx-
will be used, take the bits from programming languages, making it imation method.
the ith to the (i – pc + 1)th blocks applicable to a large range of real-life
of A and B and name them as A′ applications (see Figure 2) and letting OTHER LOW-POWER VLSI
and B′, respectively. designers control the accuracy–power TECHNIQUES FOR THE IOT
3. Compute S′ = A′ + B′ and the tradeoff at a high level. Technical In addition to approximate computing,
possible carry-out, Cout. details can be found elsewhere.11 there are many effective techniques

JUNE 2017 29
VLSI FOR THE INTERNET OF THINGS

TABLE 2. Fibonacci sequence error of the first 40 terms computed


using approximate addition under different configurations.
Term no. pc = 2 pc = 3 pc = 4

1–13 0 0 0

14 –0.00984 0 0

15 –0.01317 0 0
for power and energy minimization.
16 –0.01691 0 0 These include clock gating, transistor
17 –0.01858 0 0 sizing, glitch and path balancing, tech-
nology mapping, temperature/­thermal
18 –0.01985 0 0 awareness, and dual-­threshold voltage
19 –0.02173 –0.00044 0 and input vector control, and don’t-care
condition optimization.
20 –0.02905 0.000183 0 Dynamic voltage and frequency sched-
uling. DVFS varies the supply volt-
21 –0.02625 –5.65E–05 0
age and clock frequency based on
22 –0.02732 3.49E–05 0 the computational load and deadline
requirements to provide the required
23 –0.02692 0 0
performance while minimizing the
24 –0.02707 1.33E–05 0 total amount of energy consumption.
State-of-the-art DVFS design has been
25 –0.02912 0.000404 8.24E–06 proposed for the functionalities that
26 –0.03225 0.000336 1.02E–05 are conventionally used in IoT devices.
This can be naturally integrated with
27 –0.03750 0.000362 9.44E–06 approximate FUs.
Probabilistic design. This design meth-
28 –0.03947 0.000352 9.72E–06
odology aims at not overdesigning sys-
29 –0.04118 0.000356 9.61E–06 tems for their worst-case scenario.5 It
utilizes the fact that many real-time IoT
30 –0.04205 0.000354 9.66E–06
applications do not require high perfor-
31 –0.02291 –0.000350 2.30E–06 mance and can tolerate a certain degree
of faultiness. Probabilistic design uses
32 –0.02267 –0.000590 –8.51E–06
prior or posterior execution information
33 –0.02276 –0.000500 –4.38E–06 and takes advantages of the unique fea-
tures of IoT devices’ functionalities to
34 –0.02273 –0.000530 –5.96E–06 relax overly rigid hardware or software
35 –0.02274 –0.000520 –5.36E–06 design implementations.

36 –0.02274 –0.000520 –5.59E–06 IOT SECURITY CHALLENGES


Approximate computing can also help
37 –0.03280 –0.000100 1.05E–06
address several IoT security challenges.
38 –0.03621 –0.000460 –5.53E–06 Because IoT devices collect, process,
and exchange massive amounts of
39 –0.04003 –0.000640 –3.02E–06
data that could be confidential or pri-
40 –0.04174 –0.000770 –3.98E–06 vacy sensitive, they must be robust
to potentially malicious attacks and
design vulnerabilities, such as the
injection of Trojans and other malware
to meet the low-power demands of IoT leakage power, and short-circuit into software and hardware, unso-
systems. power. In today’s VLSI technology, phisticated electronic design automa-
Circuit-level low-power techniques. dynamic and leakage power are the tion (EDA) tools,3 and untrusted sup-
At the circuit level, three types of dominant sources. Researchers have ply chains. Although there are many
power are dissipated: dynamic power, proposed various circuit-level methods well-designed security protocols

30 COMPUTER  W W W.CO M P U T E R .O R G /CO M P U T E R


based on modern cryptography, they Side-channel attacks can be very effec- distributed IPs. Digital fingerprinting
are too computationally expensive tive against IoT devices, and they are addresses this problem by embedding
and power hungry to be suitable for easy to implement and hard to detect the IP user’s information together
IoT devices’ low-power design require- and prevent. Because they target vul- with the IP author’s signature. Finger-
ments. As a result, many IoT devices nerabilities in the hardware or soft- printed IPs can thus be viewed as mul-
such as implantable medical devices ware implementation instead of the tiple, distinct watermarked IPs. The
have no data protection. algorithms or protocols, theoretically goal is to identify each copy of the IP
Active research on lightweight cryp­ proved secure algorithms or protocols and protect honest IP users.
tography aims to deliver affordable but can also become vulnerable.
weak security (for example, with short Circuit obfuscation. While digital water­
cryptographic keys) to IoT devices12,13 Reverse engineering. RE is the inva- marking and fingerprinting techniques
as well as to hardware security primi- sive process of extracting IP from an can deter IP piracy, they do not prevent
tives such as silicon physical unclonable IoT device and reproducing it based on RE. Circuit obfuscation modifies the
functions (PUFs) and hardware-based the obtained information with little or design and implementation of a circuit
random-­ number generators used for no R&D. These low-cost illegitimate to make it difficult to interpret the lay-
authentication. These approaches are products can be sold at a much lower out and thus increases the cost and
much more energy efficient than clas- price, giving them an unfair compet- complexity of RE attacks. There are
sic cryptographic solutions. However, itive edge against the authentic prod- two types of obfuscation techniques.
they add security as a nonfunctional ucts. Moreover, after extracting the Logic encryption inserts additional key
feature to the system and therefore IoT device’s high-level functionality, gates that have the secret key values
need hardware or software support. the attacker can redesign the device to as part of their input signals; the key
avoid copyright infringement or insert gates and thus the entire circuit will
Hardware-based attacks a hardware Trojan for malicious pur- malfunction with incorrect key val-
As the fundamental element in the poses. RE attacks can be very effective ues. Circuit camouflage replaces the
IoT, hardware is becoming an attrac- against IoT devices because many are original logic gates with configurable
tive new attack surface using various simple and do not use the latest semi- logic cells. These cells can be con-
techniques. conductor technology (which raises figured to perform as different logic
the cost of RE tools). gates, but the difference between con-
Hardware Trojan. This is a piece of figurations is too small to be observed
circuit implanted in a system design or Countermeasures by existing RE tools.
modified from the original design for The hardware design community has
malicious purposes. As simple as sev- responded to these challenges with sev- Physical unclonable function. A PUF
eral logic gates, a hardware Trojan can eral countermeasures. is a small piece of circuitry embedded
cause severe damage by, for example, in the design that extracts a silicon
altering or disabling certain FUs, leak- IP watermarking and digital finger- chip’s fabrication variation and uses
ing sensitive information, or shorten- printing. IP watermarking involves this intrinsic physical feature for secu-
ing the lifetime of IoT devices. embedding the IP author’s signature rity applications. A PUF can be effective
in a circuit to claim IP rights and is for secret key storage, random num-
Side-channel attack. This classic non- used to detect and prevent IP piracy. ber generation, chip authentication,
invasive and passive attack monitors, A carefully designed watermark can IP protection, and anti-counterfeiting.
measures, and analyzes a system’s provide high confidence of IP author- While a promising hardware security
physical characteristics leaking from ship, incurs low design overhead, and primitive for IoT devices, its usability
side channels when the system is run- is resilient against various watermark is limited due to unreliability under
ning. These characteristics include removal and modification techniques. different operating environments such
timing, current, voltage, electromag- However, a watermark remains the as supply voltage, temperature, and
netic radiation, power consumption, same for all copies of a given IP and humidity as well as circuit aging.
and optical or acoustic information. cannot trace the source of illegally Therefore, researchers have proposed

JUNE 2017 31
VLSI FOR THE INTERNET OF THINGS

information without changing the


31 30 22 23 0 rest of the 32 – p bits. In the above
example, setting the last 10 bits to 0s
will make the value 3.1413574, only
Exponent Mantissa
Sign bit a 0.0074 percent decrease. The error
IEEE 754 single-precision format introduced by any changes to the last
p bits will be less than 2p – 24 relative
31 30 22 23 p p−1 0 to the original value.

Approximate computing with


Exponent Mantissa Security bits information embedding
Sign bit
Given two real numbers A and B, we
Approximate format with security embedding
rewrite them in the approximate data
format: A = A′ ⊕ K A and B = B′ ⊕ K B,
FIGURE 3. IEEE 754 single-precision floating-point format for 32-bit data. The last p where A′ and B′ are identical to A and B
bits can be used as security bits to embed information without changing the rest of the except that the last p mantissa bits are
32 – p bits. replaced by 0s; K A and K B are the last
p bits of A and B; and ⊕ is the bitwise
XOR for the last p bits. For any binary
various means to enhance PUF robust- extra hardware primitives and corre- arithmetic operation A ⊕ B, we pro-
ness and stability. sponding systems. pose the following method to perform
approximate computing and informa-
APPROXIMATE ARITHMETIC Floating-point format with tion embedding simultaneously:
COMPUTING–BASED security embedding
INFORMATION HIDING IEEE 754 specifies an approximate 1. Rewrite A and B in the approxi-
To minimize the power cost of IoT single-­ precision floating-point format mate data format.
devices while still providing a practi- for 32-bit data, which consists of 1 sign 2. Compute A′ ⊕ B′ and rewrite it
cal security solution, we introduce a bit, 8 exponent bits, and 23 mantissa as O′ ⊕ KO′.
novel approximate computing–based bits,14 as the top of Figure 3 shows. 3. Generate a p-bit secret key KS to
approach to embed information for (Double precision is very expensive and be embedded.
authentication and other security-­ seldom used in IoT devices.) The value 4. Return O′ ⊕ KS as the result of
related applications. The idea is of an IEEE 754 number is computed as A ⊕ B.
inspired by data segmentation, with sign × mantissa × 2exponent, where the
operands divided into most significant sign is 1 or –1 if the leading bit is 0 or For example, if A = 3.14159 and B =
bits (MSBs) and least significant bits 1, respectively; the mantissa is a real 12.31, A × B will be 38.6729729. Figure 4
(LSBs). In approximate computing, value between 1.0 and 2.0, with a frac- shows the result of following the above
MSBs are used for precise operation tional part represented in binary for- scheme with p = 10, in which KS is gen-
and should be preserved, but LSBs can mat; and the exponent equals the 8 bits erated using the simple bitwise XOR of
either be ignored or replaced by simple in the middle minus 127. For example, K A,KB,KO′ and a random key = 01010101.
operations such as OR. Our approach 0,10000000,10010010000111111010000 The output value is 38.67124, only
hides information into LSBs such is 1 × 1.570795 … × 2, which is roughly 0.00448 percent less than the accurate
that it does not affect approximate 3.14159 in decimal format. result of 38.6729729.
computation but can be recovered for Obviously, the LSBs in the man-
security purposes. The implementa- tissa will have little impact on the Information hiding for
tion requires slight modification to value. We thus propose using the security applications
the arithmetic unit—for example, the last p bits as security bits, as shown KS can be as simple as a constant or
adder or multiplier—without building in the bottom of Figure 3, to embed some function of K A, K B, KO′ and Key.

32 COMPUTER  W W W.CO M P U T E R .O R G /CO M P U T E R


KA
A =0,10000000,10010010000111111010000
A′ =0,10000000,10010010000110000000000
KB
B =0,10000010,10001001111010111000011
B′ =0,10000010,10001001111010000000000
KO′
A′ x B′ =0,10000100,00110101010110011001111
O′ =0,10000100,00110101010110000000000
In general, we can write this as K S = authorized parties will reveal the hid- KS =KA + KB + KO + Key=1110001001
F(K A, K B, KO′, Key). KS can be generated den information to verify the water- Output: 0,10000100,00110101010111110001001
and used for IP watermarking, dig- mark and fingerprint or decrypt the
ital fingerprinting, and lightweight encrypted results. An insider attacker FIGURE 4. Example of approximate com-
encryption. or someone who has managed to gain puting with information hiding. If p = 10, a
this permission can further alter or secret key KS is generated using the simple
IP watermarking. During IP design forge the hidden information. Such bitwise XOR of KA,KB,KO′ and a random
and implementation, either the IP attacks can be easily prevented by key = 01010101. The output value is
owner’s digital signature can be used enhancing the proposed method with 38.67124, only 0.00448 percent less
as the Key or operand values can be additional cost. For example, the sim- than the accurate result of 38.6729729.
selected to enable the above proposed ple XOR operation can be replaced by
steps—that is, a mechanism such as a one-way hash function, informa-
returning A′ ⊕ B′ directly (and skip- tion can be hidden in multiple oper- ABI Research, 9 May 2013; www.abi
ping steps 3 and 4) unless the oper- ands and at different computational research.com/press/more-than-30
ands match the given values when the stages during a complex operation, and -billion-devices-will-wirelessly-conne.
watermark is embedded or a water- intrinsic hardware features such PUFs 2. T. Danova, “Morgan Stanley: 75 Bil-
mark verification signal is activated. and secure memory can be utilized to lion Devices Will Be Connected to the
Checking the output error then reveals secure Key and KS. Internet of Things by 2020,” Business
the watermark. Insider, 2 Oct. 2013; www.business

T
insider.com/75-billion-devices-will
Digital fingerprinting. Similar to IP he main advantage of our pro- -be-connected-to-the-internet-by
watermarking, each device’s unique posed approximate computing–­ -2020-2013-10.
fingerprint can be embedded as the based information-­hiding method 3. G. Qu and L. Yuan, “Design Things
error value in the LSBs. For the same is its low implementation cost with for the Internet of Things: An EDA
verification inputs or operands, differ- guaranteed results: it utilizes the energy Perspective,” Proc. IEEE/ACM Int’l
ent Key values and F()s can be used to efficiency of approximate computing Conf. Computer-Aided Design (ICCAD
distinguish individual devices. and introduces tolerable error. The hid- 14), 2014, pp. 411–416.
den information can be verified easily 4. H. Jayakumar et al., “Powering the
Lightweight encryption. As the by disabling the approximate comput- Internet of Things,” Proc. Int’l Symp.
result will be encrypted, the entire ing. This security primitive can pro- Low Power Electronics and Design
32-bit space can be used for informa- vide lightweight security for multiple (ISLPED 14), 2014, pp. 375–380.
tion hiding. For example, an encryp- security applications such as device 5. S. Hua, G. Qu, and S.S. Bhattacha-
tion key up to 32 bits can be generated authentication and IP protection. ryya, “Energy Reduction Techniques
from F() based on the values of oper- for Multimedia Applications with
ands and some given Key values to ACKNOWLEDGMENTS Tolerance to Deadline Misses,” Proc.
encrypt the (approximated) computa- This work was supported in part by AFOSR 40th Ann. Design Automation Conf.
tional result—for example, by the effi- MURI under award number FA9550-14-1- (DAC 03), 2003, pp. 131–136.
cient bitwise XOR operation. This is 0351. Yongqiang Lyu was supported in 6. S.H. Kim, S. Mukhopadhyay, and M.
symmetric encryption and the result part by  Tsinghua University Initiative Wolf, “System-Level Energy Opti-
can be decrypted easily once the key Scientific Research Program and Chinese mization for Error-Tolerant Image
is available. National Key Research and Development Compression,” IEEE Embedded System
There are two potential threats Program 2016YFB1000102. Letters, vol. 2, no. 3, 2010, pp. 81–84.
to this information-hiding protocol. 7. S.H. Kim, S. Mukhopadhyay, and
First, an attacker can study the values REFERENCES M. Wolf, “Modeling and Analysis of
of the operands and results to reveal 1. “More Than 30 Billion Devices Will Image Dependence and Its Implica-
the Key value, the secret KS, and the Wirelessly Connect to the Internet of tions for Energy Savings in Error Tol-
function that generates KS. Second, Everything in 2020,” press release, erant Image Processing,” IEEE Trans.

JUNE 2017 33
VLSI FOR THE INTERNET OF THINGS

ABOUT THE AUTHORS


MINGZE GAO is a PhD student in the Department of Electrical and Computer
Engineering at the University of Maryland, College Park. His research interests
include hardware security, low-power design, and approximate computing.
Contact him at mgao1@umd.edu.

Computer-Aided Design of Integrated QIAN WANG is a PhD student in the Maryland Embedded Systems and Hard-
Circuits and Systems, vol. 30, no. 8, ware Security (MeshSec) lab at the University of Maryland, College Park. Her
2011, pp. 1163–1172. research interests include embedded systems and hardware security. Contact
8. Q. Xu, T. Mytkowicz, and N.S. Kim, her at qwang126@umd.edu.
“Approximate Computing: A Sur-
vey,” IEEE Design & Test, vol. 33, no. 1, MD TANVIR ARAFIN is a PhD student in the Department of Electrical and Com-
2016, pp. 8–22. puter Engineering at the University of Maryland, College Park. His research
9. R. Ye et al., “On Reconfiguration- interests include semiconductor physics, integrated circuits, and embedded
Oriented Approximate Adder Design security of microelectronic devices. Contact him at marafin@umd.edu.
and Its Application,” Proc. Int’l Conf.
Computer-Aided Design (ICCAD 13), YONGQIANG LYU is an associate professor at the Research Institute of Information
2013, pp. 48–54. Technology at Tsinghua University. His research focuses on the hardware–software
10. S. Hashemi, R.I. Bahar, and S. Reda, fusion architecture in advanced computing systems and high-­performance net-
“DRUM: A Dynamic Range Unbiased works. Lyu received a PhD in computer science from Tsinghua University. He is
Multiplier for Approximate Appli- a member of IEEE, ACM, and the China Computer Federation. Contact him at
cations,” Proc. IEEE/ACM Int’l Conf. luyq@tsinghua.edu.cn.
Computer-Aided Design (ICCAD 15),
2015, pp. 418–425. GANG QU is a professor in the Department of Electrical and Computer Engi-
11. M. Gao et al., “A Novel Data Format neering and at the Institute for Systems Research at the University of Maryland,
for Approximate Arithmetic Com- College Park, where he also directs the MeshSec and Wireless Sensor labs. His
puting,” Proc. 22nd Asia and South research interests include optimization and combinatorial problems and their
Pacific Design Automation Conf. application to very large-scale integration and computer-aided design, wire-
(ASP-DAC 17), 2017; doi:10.1109 less sensor networks, bioinformatics, cybersecurity, and hardware security and
/ASPDAC.2017.7858354. its impact on Internet of Things and embedded systems. Qu received a PhD in
12. M. Katagi and S. Moriai, “Light- computer science from the University of California, Los Angeles. He is a Senior
weight Cryptography for the Internet Member of IEEE. Contact him at gangqu@umd.edu.
of Things,” Sony Corp., 2008; www
.iab.org/wp-content/IAB-uploads
/2011/03/Kaftan.pdf.
13. M.T. Arafin, M. Gao, and G. Qu,
july • august 2016
IEEE INTERNET COMPUTING

“VOLtA: Voltage Over-Scaling Based


march • april 2016
IEEE INTERNET COMPUTING

may • june 2016


IEEE INTERNET COMPUTING

Lightweight Authentication for IoT


nov ember • december 2015
IEEE INTERNET COMPUTING

jUly • aUGUsT 2016

january • february 2016


IEEE INTERNET COMPUTING
MaRCh • aPRIl 2016

May • jUNE 2016

Applications,” Proc. 22nd Asia and


NOvEMbER • DECEMbER 2015

jaNUaRy • fEbRUaRy 2016

South Pacific Design Automation Conf.


MEasURING ThE INTERNET
ExPlORING TOMORROw’s INTERNET

(ASP-DAC 17), 2017; doi:10.1109


ThE INTERNET Of YOU

ClOUd sTORaGE

INTERNET ECONOMICs

/ASPDAC.2017.7858345.
VOl. 20, NO. 1
vOl. 19, NO. 6

14. IEEE Std. 754-2008, IEEE Standard for


VOl. 20, NO. 3
VOl. 20, NO. 2

VOl. 20, NO. 4

www.COMPUTER.ORG/INTERNET/
www.COMPUTER.ORG/INTERNET/

Floating-Point Arithmetic, IEEE, 2008.


www.COMPUTER.ORG/INTERNET/
www.COMPUTER.ORG/INTERNET/

www.COMPUTER.ORG/INTERNET/

IC-20-01-c1 Cover-1 December 7, 2015 1:45 PM

IC-19-06-c1 Cover-1 October 9, 2015 3:26 PM

IC-20-03-c1 Cover-1 April 13, 2016 8:45 PM


IC-20-02-c1 Cover-1 February 11, 2016 10:30 PM

Want to know more about the Internet?


This magazine covers all aspects of Internet computing,
Read your subscriptions from programming and standards to security and networking.
through the myCS
publications portal at
http://mycs.computer.org
www.computer.org/internet
34 COMPUTER  W W W.CO M P U T E R .O R G /CO M P U T E R

Das könnte Ihnen auch gefallen